Spaces:
Running
Running
File size: 16,886 Bytes
f492127 db7762a f492127 db7762a f492127 db7762a f492127 | 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 | """
HaramGuard โ CoordinatorAgent
================================
AISA Layer : Reasoning + Governance
Design Pattern : ReAct (Reason โ Act โ Observe) + Output Guardrails
ReAct Pattern Implementation (Required by Capstone Rubric):
This agent implements the ReAct (Reasoning-Acting-Observing) design pattern
as explicitly required by the capstone project rubric. The ReAct loop enables
iterative self-correction through:
1. REASON: Analyze the situation and compose a structured prompt
- Input: RiskResult, Decision, recent frames, feedback (if any)
- Output: Contextualized prompt for LLM
2. ACT: Execute the action (call LLM to generate action plan)
- Model: Groq API (GPT-OSS-120B or similar)
- Output: Raw JSON plan from LLM
3. OBSERVE: Validate the output using guardrails
- Run 5 validation checks (GR-C1 through GR-C5)
- If issues found: generate feedback and loop back to REASON
- If valid: return the plan
The loop continues up to MAX_REACT_ITERS (3) times, ensuring the agent
can self-correct errors in its reasoning or output format.
Responsibilities:
- Called on ALL decisions (P0/P1/P2) โ not just critical emergencies
- Implements a ReAct loop (max 3 iterations):
Reason : analyse the crowd situation and compose a prompt
Act : call LLM (Groq) to generate a structured action plan
Observe : run 6 guardrails to validate the output
โ if validation fails, feed back issues and Reason again
- Guardrails:
GR-C1: Required fields check
GR-C2: Valid threat level (CRITICAL/HIGH/MEDIUM/LOW only)
GR-C3: Confidence score in [0, 1]
GR-C4: Consistency check (low risk score โ CRITICAL threat)
GR-C5: Arabic alert fallback if empty
GR-C6: selected_gates must be a non-empty list
"""
import json
import numpy as np
from typing import Optional, Tuple
from groq import Groq
from openai import OpenAI
from core.models import RiskResult, Decision
class CoordinatorAgent:
REQUIRED_FIELDS = {
'threat_level', 'executive_summary', 'selected_gates',
'immediate_actions', 'actions_justification',
'arabic_alert', 'confidence_score'
}
VALID_THREATS = {'CRITICAL', 'HIGH', 'MEDIUM', 'LOW'}
MAX_REACT_ITERS = 3 # ReAct: maximum reasoning iterations
# Real gate list โ injected into every LLM prompt so the agent can choose
HARAM_GATES = [
'ุจุงุจ ุงูู
ูู ุนุจุฏุงูุนุฒูุฒ', # South, main entrance, highest traffic
'ุจุงุจ ุงูู
ูู ููุฏ', # North, large capacity
'ุจุงุจ ุงูุณูุงู
', # East, historic, medium traffic
'ุจุงุจ ุงููุชุญ', # West, medium capacity
'ุจุงุจ ุงูุนู
ุฑุฉ', # West, Umrah pilgrims
'ุจุงุจ ุงูู
ูู ุนุจุฏุงููู', # South-West, high capacity
'ุจุงุจ ุงูุตูุง', # East, leads to Safa-Marwa
'ุจุงุจ ุนูู', # North-East, smaller gate
'ุจุงุจ ุงูุฒูุงุฏุฉ', # North, overflow gate
'ุจุงุจ ุงูู
ุฑูุฉ', # East, leads to Marwa
]
def __init__(self, groq_api_key: str):
self.name = 'CoordinatorAgent'
self.aisa_layer = 'Reasoning + Governance (ReAct)'
self._groq_client = Groq(api_key=groq_api_key)
self._active_backend = 'groq'
self._active_model = 'llama-3.3-70b-versatile'
print(f'๐ง [CoordinatorAgent] Ready โ backend=groq model={self._active_model} | ReAct loop')
# โโ Guardrails โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def _validate(self, plan: dict, risk_score: float) -> Tuple[dict, list]:
"""
Validate and sanitise model output.
Returns (corrected_plan, list_of_issues).
Issues list is empty when the plan is fully valid.
"""
issues = []
if not isinstance(plan, dict):
plan = {}
issues.append('GR_C1_invalid_json_object')
# GR-C1: Required fields
for field_name in self.REQUIRED_FIELDS:
if field_name not in plan:
plan[field_name] = 'N/A'
issues.append(f'GR_C1_missing:{field_name}')
# GR-C2: Valid threat level
tl = str(plan.get('threat_level', '')).upper()
if tl not in self.VALID_THREATS:
issues.append(f'GR_C2_invalid_threat:{tl}->HIGH')
plan['threat_level'] = 'HIGH'
else:
plan['threat_level'] = tl
# GR-C3: Confidence in [0, 1]
cs = plan.get('confidence_score', 0)
if not isinstance(cs, (int, float)) or not (0 <= cs <= 1):
issues.append(f'GR_C3_invalid_confidence:{cs}->0.5')
plan['confidence_score'] = 0.5
# GR-C4: Consistency โ threat_level must match risk_score thresholds
# risk > 0.80 (density_pct > 80%) โ HIGH
# risk > 0.20 (density_pct > 20%) โ MEDIUM
# risk <= 0.20 (density_pct <= 20%) โ LOW
expected = 'HIGH' if risk_score > 0.80 else 'MEDIUM' if risk_score > 0.20 else 'LOW'
tl_current = plan['threat_level']
if tl_current != expected:
issues.append(f'GR_C4_threat_corrected:{tl_current}->{expected}')
plan['threat_level'] = expected
# GR-C5: Arabic alert fallback
if not str(plan.get('arabic_alert', '')).strip():
plan['arabic_alert'] = (
'ุชูุจูู ุฃู
ูู: ููุฑุฌู ู
ุฑุงูุจุฉ ููุงุท ุงูุชุฌู
ุน ูุงุชุฎุงุฐ ุงูุฅุฌุฑุงุกุงุช ุงูููุงุฆูุฉ ุงููุงุฒู
ุฉ.'
)
issues.append('GR_C5_arabic_fallback')
# GR-C1 extra: immediate_actions must be a non-empty list
ia = plan.get('immediate_actions', [])
if not isinstance(ia, list) or not ia:
plan['immediate_actions'] = ['ุฒูุงุฏุฉ ุงูู
ุฑุงูุจุฉ ุงูู
ูุฏุงููุฉ', 'ุฅุฑุณุงู ูุญุฏุงุช ุฅูู ููุทุฉ ุงูุงุฒุฏุญุงู
']
issues.append('GR_C1_immediate_actions_fixed')
# GR-C6: selected_gates โ count enforced by threat level
# LOW/P2 โ 0 gates (no action needed)
# MEDIUM โ exactly 1 gate
# HIGH โ exactly 2 gates
# CRITICALโ exactly 2 gates (same as HIGH)
tl_now = plan.get('threat_level', 'LOW')
sg = plan.get('selected_gates', [])
if not isinstance(sg, list):
sg = []
if tl_now == 'LOW':
plan['selected_gates'] = [] # no action, no gates shown
elif tl_now == 'MEDIUM':
if not sg:
plan['selected_gates'] = ['ุจุงุจ ุงูู
ูู ุนุจุฏุงูุนุฒูุฒ']
issues.append('GR_C6_medium_fallback')
else:
plan['selected_gates'] = sg[:1] # cap at 1
else: # HIGH or CRITICAL
if len(sg) < 2:
fallback = ['ุจุงุจ ุงูู
ูู ุนุจุฏุงูุนุฒูุฒ', 'ุจุงุจ ุงูุณูุงู
']
plan['selected_gates'] = (sg + fallback)[:2]
issues.append('GR_C6_high_padded')
else:
plan['selected_gates'] = sg[:2] # cap at 2
plan['_guardrail_issues'] = issues
return plan, issues
# โโ ReAct helpers โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def _build_prompt(
self,
rr: RiskResult,
decision: Decision,
recent_frames: list,
feedback: str = ''
) -> str:
"""
Reason step: compose the model prompt.
If feedback is provided (from a previous failed Observe step),
include it so the model can correct itself.
"""
avg_p = np.mean([f.person_count for f in recent_frames]) if recent_frames else 0
cur_count = recent_frames[-1].person_count if recent_frames else 0
gates_list = '\n'.join(
f'{i+1}. {g}' for i, g in enumerate(self.HARAM_GATES)
)
base = (
'ุฃูุช ู
ูุณู ูุธุงู
ุญุงุฑุณ ุงูุญุฑู
ูุฅุฏุงุฑุฉ ุณูุงู
ุฉ ุงูุญุดูุฏ ูู ุงูู
ุณุฌุฏ ุงูุญุฑุงู
.\n'
'ู
ูู
ุชู: ุฅูุชุงุฌ ุฎุทุฉ ุชุดุบูููุฉ ูุงุถุญุฉ ูู
ูุฌุฒุฉ ููู
ุดุบููู ุจูุงุกู ุนูู ุจูุงูุงุช ุงูุญุดูุฏ ุงูุญุงููุฉ.\n\n'
f'ู
ุณุชูู ุงูุฎุทุฑ : {rr.risk_level} (ุงูุฏุฑุฌุฉ: {rr.risk_score:.3f})\n'
f'ุงูุฃููููุฉ : {decision.priority} (P0=ุทุงุฑุฆ ุญุฑุฌุ P1=ุชุญุฐูุฑ ููุงุฆูุ P2=ู
ุฑุงูุจุฉ ุฑูุชูููุฉ)\n'
f'ุงุชุฌุงู ุงูุญุดูุฏ : {rr.trend}\n'
f'ุงูุนุฏุฏ ุงูุญุงูู : {cur_count} ุดุฎุต | ุงูู
ุชูุณุท (ุขุฎุฑ 30 ุฅุทุงุฑ): {avg_p:.0f} | ุงูุฐุฑูุฉ: {rr.window_max}\n\n'
'ุงูุจูุงุจุงุช ุงูู
ุชุงุญุฉ ูู ุงูู
ุณุฌุฏ ุงูุญุฑุงู
โ ุงุฎุชุฑ ุงูุฃูุณุจ ู
ููุง:\n'
f'{gates_list}\n\n'
'ุฅุฑุดุงุฏุงุช ุงุฎุชูุงุฑ ุงูุจูุงุจุงุช:\n'
' P0 (ุทุงุฑุฆ ุญุฑุฌ): ุงูุชุญ ุจูุงุจุงุช ุงูุฅุฎูุงุก ุนุงููุฉ ุงูุณุนุฉ + ุฃุบูู ุงูู
ุฏุงุฎู ุงูู
ุฒุฏุญู
ุฉ\n'
' P1 (ุชุญุฐูุฑ) : ูุนูู ุงููุงูุชุงุช ุงูุฅุฑุดุงุฏูุฉ ูุญู ุงูุจูุงุจุงุช ุงูุฃูู ุงุฒุฏุญุงู
ุงู\n'
' P2 (ุฑูุชููู) : ุฑุงูุจ ุงูุจูุงุจุงุช ุงูุฃูุซุฑ ุญุฑูุฉ ููุท\n\n'
)
if feedback:
base += (
'ุชุตุญูุญ ู
ุทููุจ โ ุงูุฅุฌุงุจุฉ ุงูุณุงุจูุฉ ุจูุง ู
ุดุงูู:\n'
f'{feedback}\n'
'ุตุญุญ ุฌู
ูุน ุงูู
ุดุงูู ูุฃุนุฏ ุงูุฅุฌุงุจุฉ.\n\n'
)
base += (
'ููุงุนุฏ ุงูุฅุฎุฑุงุฌ ุงูุตุงุฑู
ุฉ:\n'
'- ุฃุฌุจ ููุท ุจู JSON ุฎุงู
. ุจุฏูู markdown. ุจุฏูู backticks. ุจุฏูู ุฃู ูุต ูุจูู ุฃู ุจุนุฏู.\n'
'- ุฌู
ูุน ุญููู ุงููุต ูุฌุจ ุฃู ุชููู ุจุงููุบุฉ ุงูุนุฑุจูุฉ ุงููุตุญู ุงูุฑุณู
ูุฉ ุญุตุฑุงู. ูุง ุชุณุชุฎุฏู
ุงูุนุงู
ูุฉ ุฃู ุงูููุฌุงุช ุงูู
ุญููุฉ ู
ุทููุงู. ุงุณุชุฎุฏู
ุฃุณููุจุงู ู
ูููุงู ุฑุณู
ูุงู ูููู ุจุฅุฏุงุฑุฉ ุงูุญุฑู
ุงูู
ูู.\n'
'- ู
ู
ููุน ุงุณุชุฎุฏุงู
ุตูุบุฉ ุงูุฃู
ุฑ ุงูู
ุจุงุดุฑ (ุฑุงูุจูุงุ ุชุฃูุฏูุงุ ุงูุชุญูุง). ุงุณุชุฎุฏู
ุฏุงุฆู
ุงู ุตูุบุฉ ููุฑุฌู / ููุทูุจ / ูููุตู. ู
ุซุงู: ููุฑุฌู ู
ุฑุงูุจุฉ... ูููุณ ุฑุงูุจูุง...\n'
'- ูุง ุชุณุชุฎุฏู
ุนูุงู
ุงุช ุงูุงูุชุจุงุณ ุงูู
ุฒุฏูุฌุฉ (") ุฏุงุฎู ููู
ุงููุต ุงูุนุฑุจู.\n'
'- ูุง ุชุถุน ุฃุณุทุฑุงู ุฌุฏูุฏุฉ (\\n) ุฏุงุฎู ููู
ุงููุตูุต ูู JSON.\n\n'
'ููุงุนุฏ selected_gates โ ุฅูุฒุงู
ูุฉ ุจุญุณุจ ู
ุณุชูู ุงูุฎุทุฑ:\n'
' * LOW (P2 ุฑูุชููู) : ูุงุฆู
ุฉ ูุงุฑุบุฉ [] โ ูุง ุญุงุฌุฉ ูุฃู ุฅุฌุฑุงุก ุนูู ุงูุจูุงุจุงุช.\n'
' * MEDIUM (P1 ุชุญุฐูุฑ): ุจูุงุจุฉ ูุงุญุฏุฉ ููุท โ ุงูุฃูุณุจ ููุชูุฌูู ุงูููุงุฆู.\n'
' * HIGH/CRITICAL (P0): ุจูุงุจุชุงู ููุท โ ุจูุงุจุฉ ุฅุฎูุงุก + ุจูุงุจุฉ ุชุญููู.\n'
' * ุงุณุชุฎุฏู
ุงูุฃุณู
ุงุก ุงูุนุฑุจูุฉ ุงูุฏูููุฉ ู
ู ุงููุงุฆู
ุฉ ุฃุนูุงู ุญุฑูุงู ุจุญุฑู.\n'
' * ู
ู
ููุน ู
ูุนุงู ุจุงุชุงู ุฅุฏุฑุงุฌ ุฃูุซุฑ ู
ู ุจูุงุจุชูู ูู ุฃู ุญุงู โ ูุฐุง ุฎุทุฃ ูุงุฏุญ.\n\n'
'- immediate_actions: 3 ุฅูู 5 ุฅุฌุฑุงุกุงุช ุนุฑุจูุฉ ูุตูุฑุฉ ุชุฐูุฑ ุงูุจูุงุจุงุช ุงูู
ุฎุชุงุฑุฉ ุจุงูุงุณู
.\n'
'- actions_justification: ุฃูู ู
ู 40 ููู
ุฉ. ุงุดุฑุญ ูู
ุงุฐุง ูุฐู ุงูุจูุงุจุงุช ุจุงูุฐุงุช.\n'
'- arabic_alert: ุฃูู ู
ู 15 ููู
ุฉ. ุชูุฌูู ุฑุณู
ู ุจุงููุตุญู ู
ูุฌูู ูู
ูุธูู ุงูุฃู
ู ูุงูู
ุดุบููู ุญุตุฑุงู. ุงุณุชุฎุฏู
ุตูุบุฉ ููุฑุฌู/ููุทูุจ ููุท. ู
ุซุงู: ููุฑุฌู ุชูุฌูู ุงูุญุดูุฏ ูุญู ุจุงุจ ุงูู
ูู ููุฏ ูู
ุฑุงูุจุฉ ููุงุท ุงูุชุฌู
ุน.\n'
'- executive_summary: ุฃูู ู
ู 20 ููู
ุฉ. ู
ูุฎุต ุงูู
ููู ููููุงุฏุฉ.\n\n'
'ุฃุนุฏ ุจุงูุถุจุท ูุฐุง JSON (ุจุฏูู ุญููู ุฅุถุงููุฉุ ุจุฏูู ูุต ุฅุถุงูู):\n'
'{\n'
' "threat_level": "HIGH",\n'
' "executive_summary": "...",\n'
' "selected_gates": ["ุจุงุจ ..."],\n'
' "immediate_actions": ["..."],\n'
' "actions_justification": "...",\n'
' "arabic_alert": "...",\n'
' "confidence_score": 0.85\n'
'}'
)
return base
def _llm_call(self, prompt: str) -> Optional[dict]:
"""Act step: call Groq LLM, parse JSON response."""
raw_text = ''
try:
resp = self._groq_client.chat.completions.create(
model=self._active_model,
messages=[{'role': 'user', 'content': prompt}],
max_tokens=1200,
temperature=0.2,
)
raw_text = (resp.choices[0].message.content or '').strip()
# โโ Parse JSON โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
try:
return json.loads(raw_text)
except json.JSONDecodeError:
pass
# Brace-counting extractor (handles truncated responses)
start = raw_text.find('{')
if start != -1:
fragment = raw_text[start:]
depth = 0
best_end = -1
for i, ch in enumerate(fragment):
if ch == '{':
depth += 1
elif ch == '}':
depth -= 1
if depth == 0:
best_end = i
break
if best_end != -1:
return json.loads(fragment[:best_end + 1])
raise json.JSONDecodeError('No valid JSON block found', raw_text, 0)
except json.JSONDecodeError as e:
print(f' [CoordinatorAgent] JSON parse error: {e}')
print(f' [CoordinatorAgent] Raw LLM response (first 600 chars):\n{raw_text[:600]}')
return {}
except Exception as e:
print(f' [CoordinatorAgent] API error: {e}')
return None
# โโ Public API โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def call(self, rr: RiskResult, decision: Decision, recent_frames: list) -> Optional[dict]:
"""
ReAct loop: Reason โ Act โ Observe, up to MAX_REACT_ITERS times.
Returns validated plan dict or None on unrecoverable error.
"""
print(f'\n๐ง [CoordinatorAgent] {decision.priority} @ frame {rr.frame_id} โ ReAct loop starting...')
feedback = ''
best_plan = None
for iteration in range(1, self.MAX_REACT_ITERS + 1):
print(f' โบ ReAct iteration {iteration}/{self.MAX_REACT_ITERS}')
# โโ Reason โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
prompt = self._build_prompt(rr, decision, recent_frames, feedback)
# โโ Act โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
raw = self._llm_call(prompt)
if raw is None:
print(' [CoordinatorAgent] API unavailable โ aborting ReAct')
return best_plan # return previous best (may be None)
# โโ Observe โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
plan, issues = self._validate(raw, rr.risk_score)
best_plan = plan
if not issues:
print(f' โ
[CoordinatorAgent] Plan validated on iteration {iteration}')
plan['_react_iterations'] = iteration
plan['_llm_model'] = f'{self._active_backend}/{self._active_model}'
return plan
feedback = '; '.join(issues)
print(f' โ ๏ธ Issues found ({len(issues)}): {feedback}')
print(' [CoordinatorAgent] Max ReAct iterations reached โ returning best effort')
if best_plan:
best_plan['_react_iterations'] = self.MAX_REACT_ITERS
best_plan['_llm_model'] = f'{self._active_backend}/{self._active_model}'
return best_plan |