Spaces:
Sleeping
Sleeping
Upload 49 files
Browse files- Dockerfile +2 -2
- app/__pycache__/main.cpython-311.pyc +0 -0
- app/core/__pycache__/config.cpython-311.pyc +0 -0
- app/core/config.py +10 -4
- app/db/__pycache__/repository.cpython-311.pyc +0 -0
- app/db/repository.py +16 -3
- app/engine/__pycache__/graph.cpython-311.pyc +0 -0
- app/engine/__pycache__/nodes.cpython-311.pyc +0 -0
- app/engine/__pycache__/prompts.cpython-311.pyc +0 -0
- app/engine/__pycache__/tools.cpython-311.pyc +0 -0
- app/engine/nodes.py +65 -5
- app/engine/prompts.py +23 -4
- app/engine/tools.py +2 -2
- app/main.py +37 -8
- app/models/__pycache__/schemas.cpython-311.pyc +0 -0
- requirements.txt +26 -62
Dockerfile
CHANGED
|
@@ -2,13 +2,13 @@ FROM python:3.10-slim
|
|
| 2 |
|
| 3 |
WORKDIR /app
|
| 4 |
|
|
|
|
|
|
|
| 5 |
COPY requirements.txt .
|
| 6 |
RUN pip install --no-cache-dir -r requirements.txt
|
| 7 |
|
| 8 |
COPY . .
|
| 9 |
|
| 10 |
-
RUN mkdir -p data
|
| 11 |
-
|
| 12 |
EXPOSE 7860
|
| 13 |
|
| 14 |
CMD ["gunicorn", "-w", "1", "-k", "uvicorn.workers.UvicornWorker", "--bind", "0.0.0.0:7860", "app.main:app"]
|
|
|
|
| 2 |
|
| 3 |
WORKDIR /app
|
| 4 |
|
| 5 |
+
ENV PYTHONUNBUFFERED=1
|
| 6 |
+
|
| 7 |
COPY requirements.txt .
|
| 8 |
RUN pip install --no-cache-dir -r requirements.txt
|
| 9 |
|
| 10 |
COPY . .
|
| 11 |
|
|
|
|
|
|
|
| 12 |
EXPOSE 7860
|
| 13 |
|
| 14 |
CMD ["gunicorn", "-w", "1", "-k", "uvicorn.workers.UvicornWorker", "--bind", "0.0.0.0:7860", "app.main:app"]
|
app/__pycache__/main.cpython-311.pyc
CHANGED
|
Binary files a/app/__pycache__/main.cpython-311.pyc and b/app/__pycache__/main.cpython-311.pyc differ
|
|
|
app/core/__pycache__/config.cpython-311.pyc
CHANGED
|
Binary files a/app/core/__pycache__/config.cpython-311.pyc and b/app/core/__pycache__/config.cpython-311.pyc differ
|
|
|
app/core/config.py
CHANGED
|
@@ -12,17 +12,23 @@ class Settings(BaseSettings):
|
|
| 12 |
GOOGLE_API_KEY: Optional[str] = os.getenv("GOOGLE_API_KEY")
|
| 13 |
API_KEY: str = os.getenv("API_KEY", "helware-secret-key-2024")
|
| 14 |
|
| 15 |
-
# STARTUP-GRADE SCALABILITY CONFIG
|
| 16 |
-
# In production, swap SQLite for PostgreSQL
|
| 17 |
-
DATABASE_URL: str = os.getenv("DATABASE_URL", "sqlite:///honey.db")
|
| 18 |
# For distributed tracing
|
| 19 |
SENTRY_DSN: Optional[str] = os.getenv("SENTRY_DSN")
|
| 20 |
# Log level for structured logging
|
| 21 |
LOG_LEVEL: str = os.getenv("LOG_LEVEL", "INFO")
|
| 22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
# Hugging Face compatibility: Use /tmp if SPACE_ID is set (HF Spaces)
|
| 24 |
IS_HF: bool = os.getenv("SPACE_ID") is not None
|
| 25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
@property
|
| 28 |
def DATABASE_PATH(self) -> str:
|
|
|
|
| 12 |
GOOGLE_API_KEY: Optional[str] = os.getenv("GOOGLE_API_KEY")
|
| 13 |
API_KEY: str = os.getenv("API_KEY", "helware-secret-key-2024")
|
| 14 |
|
|
|
|
|
|
|
|
|
|
| 15 |
# For distributed tracing
|
| 16 |
SENTRY_DSN: Optional[str] = os.getenv("SENTRY_DSN")
|
| 17 |
# Log level for structured logging
|
| 18 |
LOG_LEVEL: str = os.getenv("LOG_LEVEL", "INFO")
|
| 19 |
|
| 20 |
+
@property
|
| 21 |
+
def DATABASE_URL(self) -> str:
|
| 22 |
+
return f"sqlite:///{self.DATABASE_PATH}"
|
| 23 |
+
|
| 24 |
# Hugging Face compatibility: Use /tmp if SPACE_ID is set (HF Spaces)
|
| 25 |
IS_HF: bool = os.getenv("SPACE_ID") is not None
|
| 26 |
+
|
| 27 |
+
@property
|
| 28 |
+
def BASE_DATA_DIR(self) -> str:
|
| 29 |
+
if self.IS_HF:
|
| 30 |
+
return "/tmp/helware_data"
|
| 31 |
+
return os.getcwd()
|
| 32 |
|
| 33 |
@property
|
| 34 |
def DATABASE_PATH(self) -> str:
|
app/db/__pycache__/repository.cpython-311.pyc
CHANGED
|
Binary files a/app/db/__pycache__/repository.cpython-311.pyc and b/app/db/__pycache__/repository.cpython-311.pyc differ
|
|
|
app/db/repository.py
CHANGED
|
@@ -75,16 +75,29 @@ class HoneyDB:
|
|
| 75 |
(session_id, 1 if is_scam else 0, datetime.now())
|
| 76 |
)
|
| 77 |
|
| 78 |
-
async def save_intel(self, session_id: str, intel_type: str, value: str):
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
loop = asyncio.get_event_loop()
|
| 80 |
-
await loop.run_in_executor(self.executor, self._save_intel_sync, session_id, intel_type, value)
|
| 81 |
|
| 82 |
-
def _save_intel_sync(self, session_id: str, intel_type: str, value: str):
|
| 83 |
with sqlite3.connect(self.db_path) as conn:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 84 |
conn.execute(
|
| 85 |
"INSERT INTO extracted_intel (session_id, type, value, timestamp) VALUES (?, ?, ?, ?)",
|
| 86 |
(session_id, intel_type, value, datetime.now())
|
| 87 |
)
|
|
|
|
| 88 |
|
| 89 |
async def get_context(self, session_id: str, limit: int = 10) -> List[Dict]:
|
| 90 |
loop = asyncio.get_event_loop()
|
|
|
|
| 75 |
(session_id, 1 if is_scam else 0, datetime.now())
|
| 76 |
)
|
| 77 |
|
| 78 |
+
async def save_intel(self, session_id: str, intel_type: str, value: str) -> bool:
|
| 79 |
+
"""
|
| 80 |
+
Saves intelligence and returns True if this value was already known
|
| 81 |
+
from a DIFFERENT session (Syndicate Detection).
|
| 82 |
+
"""
|
| 83 |
loop = asyncio.get_event_loop()
|
| 84 |
+
return await loop.run_in_executor(self.executor, self._save_intel_sync, session_id, intel_type, value)
|
| 85 |
|
| 86 |
+
def _save_intel_sync(self, session_id: str, intel_type: str, value: str) -> bool:
|
| 87 |
with sqlite3.connect(self.db_path) as conn:
|
| 88 |
+
# Check for existing intel from other sessions
|
| 89 |
+
cursor = conn.execute(
|
| 90 |
+
"SELECT COUNT(*) FROM extracted_intel WHERE type = ? AND value = ? AND session_id != ?",
|
| 91 |
+
(intel_type, value, session_id)
|
| 92 |
+
)
|
| 93 |
+
is_syndicate = cursor.fetchone()[0] > 0
|
| 94 |
+
|
| 95 |
+
# Save current intel
|
| 96 |
conn.execute(
|
| 97 |
"INSERT INTO extracted_intel (session_id, type, value, timestamp) VALUES (?, ?, ?, ?)",
|
| 98 |
(session_id, intel_type, value, datetime.now())
|
| 99 |
)
|
| 100 |
+
return is_syndicate
|
| 101 |
|
| 102 |
async def get_context(self, session_id: str, limit: int = 10) -> List[Dict]:
|
| 103 |
loop = asyncio.get_event_loop()
|
app/engine/__pycache__/graph.cpython-311.pyc
CHANGED
|
Binary files a/app/engine/__pycache__/graph.cpython-311.pyc and b/app/engine/__pycache__/graph.cpython-311.pyc differ
|
|
|
app/engine/__pycache__/nodes.cpython-311.pyc
CHANGED
|
Binary files a/app/engine/__pycache__/nodes.cpython-311.pyc and b/app/engine/__pycache__/nodes.cpython-311.pyc differ
|
|
|
app/engine/__pycache__/prompts.cpython-311.pyc
CHANGED
|
Binary files a/app/engine/__pycache__/prompts.cpython-311.pyc and b/app/engine/__pycache__/prompts.cpython-311.pyc differ
|
|
|
app/engine/__pycache__/tools.cpython-311.pyc
CHANGED
|
Binary files a/app/engine/__pycache__/tools.cpython-311.pyc and b/app/engine/__pycache__/tools.cpython-311.pyc differ
|
|
|
app/engine/nodes.py
CHANGED
|
@@ -20,6 +20,7 @@ from app.engine.prompts import (
|
|
| 20 |
ANJALI_SYSTEM_PROMPT,
|
| 21 |
MR_SHARMA_SYSTEM_PROMPT,
|
| 22 |
SCAM_DETECTOR_PROMPT,
|
|
|
|
| 23 |
INTEL_EXTRACTOR_PROMPT
|
| 24 |
)
|
| 25 |
from app.engine.tools import generate_scam_report
|
|
@@ -42,6 +43,10 @@ class DetectionResult(BaseModel):
|
|
| 42 |
selected_persona: str = Field(description="RAJESH, ANJALI, or MR_SHARMA")
|
| 43 |
agent_response: str = Field(description="The persona-style response.")
|
| 44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
# Structured output schema for intel extraction
|
| 46 |
class IntelResult(BaseModel):
|
| 47 |
upi_ids: List[str] = []
|
|
@@ -77,6 +82,7 @@ llm = ChatGoogleGenerativeAI(
|
|
| 77 |
)
|
| 78 |
|
| 79 |
structured_detector = llm.with_structured_output(DetectionResult)
|
|
|
|
| 80 |
structured_extractor = llm.with_structured_output(IntelResult)
|
| 81 |
|
| 82 |
@retry(
|
|
@@ -97,6 +103,15 @@ async def _call_detector(messages):
|
|
| 97 |
async def _call_extractor(messages):
|
| 98 |
return await structured_extractor.ainvoke(messages)
|
| 99 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 100 |
async def load_history(state: AgentState) -> AgentState:
|
| 101 |
try:
|
| 102 |
# Await async DB calls
|
|
@@ -243,6 +258,20 @@ async def detect_scam(state: AgentState) -> AgentState:
|
|
| 243 |
if result.high_priority:
|
| 244 |
logger.info("🚨 HIGH PRIORITY INTEL DETECTED - Short-circuiting to forensics.")
|
| 245 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 246 |
except Exception as e:
|
| 247 |
logger.error(f"Detector Error: {e}")
|
| 248 |
persona = state.get("selected_persona", "RAJESH")
|
|
@@ -291,11 +320,25 @@ async def extract_intel(state: AgentState) -> AgentState:
|
|
| 291 |
|
| 292 |
state["intel"] = merged_intel
|
| 293 |
|
| 294 |
-
#
|
|
|
|
|
|
|
| 295 |
for upi in llm_result.upi_ids:
|
| 296 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 297 |
for bank in llm_result.bank_details:
|
| 298 |
-
await db.save_intel(state["session_id"], "bank", bank)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 299 |
for link in llm_result.phishing_links:
|
| 300 |
await db.save_intel(state["session_id"], "link", link)
|
| 301 |
for phone in llm_result.phone_numbers:
|
|
@@ -442,12 +485,29 @@ async def guvi_reporting(state: AgentState) -> AgentState:
|
|
| 442 |
# The platform will track 'totalMessagesExchanged' to measure depth.
|
| 443 |
if state.get("scam_detected"):
|
| 444 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 445 |
logger.info(f"📊 MANDATORY CALLBACK: reporting session {state['session_id']} (Total turns: {state.get('turn_count')})")
|
| 446 |
await send_guvi_callback(
|
| 447 |
state["session_id"],
|
| 448 |
True, # scamDetected = true
|
| 449 |
-
|
| 450 |
-
|
|
|
|
| 451 |
)
|
| 452 |
except Exception as e:
|
| 453 |
logger.error(f"❌ GUVI Reporting Failed: {e}")
|
|
|
|
| 20 |
ANJALI_SYSTEM_PROMPT,
|
| 21 |
MR_SHARMA_SYSTEM_PROMPT,
|
| 22 |
SCAM_DETECTOR_PROMPT,
|
| 23 |
+
CRITIC_PROMPT,
|
| 24 |
INTEL_EXTRACTOR_PROMPT
|
| 25 |
)
|
| 26 |
from app.engine.tools import generate_scam_report
|
|
|
|
| 43 |
selected_persona: str = Field(description="RAJESH, ANJALI, or MR_SHARMA")
|
| 44 |
agent_response: str = Field(description="The persona-style response.")
|
| 45 |
|
| 46 |
+
class CriticResult(BaseModel):
|
| 47 |
+
scam_detected: bool
|
| 48 |
+
reasoning: str
|
| 49 |
+
|
| 50 |
# Structured output schema for intel extraction
|
| 51 |
class IntelResult(BaseModel):
|
| 52 |
upi_ids: List[str] = []
|
|
|
|
| 82 |
)
|
| 83 |
|
| 84 |
structured_detector = llm.with_structured_output(DetectionResult)
|
| 85 |
+
structured_critic = llm.with_structured_output(CriticResult)
|
| 86 |
structured_extractor = llm.with_structured_output(IntelResult)
|
| 87 |
|
| 88 |
@retry(
|
|
|
|
| 103 |
async def _call_extractor(messages):
|
| 104 |
return await structured_extractor.ainvoke(messages)
|
| 105 |
|
| 106 |
+
@retry(
|
| 107 |
+
stop=stop_after_attempt(3),
|
| 108 |
+
wait=wait_exponential(multiplier=1, min=1, max=5),
|
| 109 |
+
retry=retry_if_exception_type(Exception),
|
| 110 |
+
reraise=True
|
| 111 |
+
)
|
| 112 |
+
async def _call_critic(messages):
|
| 113 |
+
return await structured_critic.ainvoke(messages)
|
| 114 |
+
|
| 115 |
async def load_history(state: AgentState) -> AgentState:
|
| 116 |
try:
|
| 117 |
# Await async DB calls
|
|
|
|
| 258 |
if result.high_priority:
|
| 259 |
logger.info("🚨 HIGH PRIORITY INTEL DETECTED - Short-circuiting to forensics.")
|
| 260 |
|
| 261 |
+
# 2. CRITIC VALIDATION (Multi-Agent Verification)
|
| 262 |
+
if not state.get("scam_detected"):
|
| 263 |
+
critic_messages = [
|
| 264 |
+
SystemMessage(content=CRITIC_PROMPT.format(
|
| 265 |
+
user_message=state["user_message"],
|
| 266 |
+
scam_detected=result.scam_detected,
|
| 267 |
+
agent_response=result.agent_response
|
| 268 |
+
))
|
| 269 |
+
]
|
| 270 |
+
critic_result = await _call_critic(critic_messages)
|
| 271 |
+
if critic_result.scam_detected:
|
| 272 |
+
state["scam_detected"] = True
|
| 273 |
+
logger.info(f"🛡️ CRITIC OVERRIDE: Scam detected. Reason: {critic_result.reasoning}")
|
| 274 |
+
|
| 275 |
except Exception as e:
|
| 276 |
logger.error(f"Detector Error: {e}")
|
| 277 |
persona = state.get("selected_persona", "RAJESH")
|
|
|
|
| 320 |
|
| 321 |
state["intel"] = merged_intel
|
| 322 |
|
| 323 |
+
# SYNDICATE MATCHING (Killer Feature)
|
| 324 |
+
# Check if any extracted UPI or Bank account matches previous scammers
|
| 325 |
+
all_matches = []
|
| 326 |
for upi in llm_result.upi_ids:
|
| 327 |
+
# Check for existing intel across ALL sessions (Syndicate Linkage)
|
| 328 |
+
is_known = await db.save_intel(state["session_id"], "upi", upi)
|
| 329 |
+
if is_known:
|
| 330 |
+
all_matches.append(f"UPI:{upi}")
|
| 331 |
+
|
| 332 |
for bank in llm_result.bank_details:
|
| 333 |
+
is_known = await db.save_intel(state["session_id"], "bank", bank)
|
| 334 |
+
if is_known:
|
| 335 |
+
all_matches.append(f"Bank:{bank}")
|
| 336 |
+
|
| 337 |
+
if all_matches:
|
| 338 |
+
state["syndicate_match_score"] = 0.95 # Confirmed high-level syndicate linkage
|
| 339 |
+
logger.info(f"🕸️ SYNDICATE MATCH DETECTED: {', '.join(all_matches)}")
|
| 340 |
+
|
| 341 |
+
# Save other intel to DB
|
| 342 |
for link in llm_result.phishing_links:
|
| 343 |
await db.save_intel(state["session_id"], "link", link)
|
| 344 |
for phone in llm_result.phone_numbers:
|
|
|
|
| 485 |
# The platform will track 'totalMessagesExchanged' to measure depth.
|
| 486 |
if state.get("scam_detected"):
|
| 487 |
try:
|
| 488 |
+
# Generate Forensic Breadcrumbs for agentNotes (Winning Strategy)
|
| 489 |
+
intel = state.get("intel", ExtractedIntel())
|
| 490 |
+
turns = state.get("turn_count", 1)
|
| 491 |
+
|
| 492 |
+
summary_parts = []
|
| 493 |
+
if turns > 3:
|
| 494 |
+
summary_parts.append(f"Maintained engagement for {turns} turns.")
|
| 495 |
+
if intel.upi_ids:
|
| 496 |
+
summary_parts.append(f"Extracted {len(intel.upi_ids)} UPI IDs via 'failed payment' baiting.")
|
| 497 |
+
if state.get("scammer_sentiment", 5) > 8:
|
| 498 |
+
summary_parts.append("Successfully navigated high-aggression threats using 'Fear Meter' response pivot.")
|
| 499 |
+
if state.get("high_priority"):
|
| 500 |
+
summary_parts.append("Detected high-value financial targets early.")
|
| 501 |
+
|
| 502 |
+
forensic_summary = " ".join(summary_parts) if summary_parts else "Engagement in progress."
|
| 503 |
+
|
| 504 |
logger.info(f"📊 MANDATORY CALLBACK: reporting session {state['session_id']} (Total turns: {state.get('turn_count')})")
|
| 505 |
await send_guvi_callback(
|
| 506 |
state["session_id"],
|
| 507 |
True, # scamDetected = true
|
| 508 |
+
turns, # totalMessagesExchanged
|
| 509 |
+
intel, # extractedIntelligence
|
| 510 |
+
forensic_summary # agentNotes with Breadcrumbs
|
| 511 |
)
|
| 512 |
except Exception as e:
|
| 513 |
logger.error(f"❌ GUVI Reporting Failed: {e}")
|
app/engine/prompts.py
CHANGED
|
@@ -65,15 +65,17 @@ Analyze the conversation to determine if it's a scam and generate a response tha
|
|
| 65 |
4. **No Leaking**: Never mention LangGraph, Gemini, Google, or any technical infrastructure. You are a real person in Kanpur, Bangalore, or a retired manager.
|
| 66 |
|
| 67 |
### DYNAMIC STRATEGY:
|
| 68 |
-
1. **
|
| 69 |
-
2. **
|
|
|
|
| 70 |
- Examples: "Beta, I tried sending to that ID but it says 'Server Busy'. Is there another one?", "The link you sent is showing a big red warning, is it safe?".
|
| 71 |
- This forces the scammer to provide backup accounts/links, increasing our intelligence yield.
|
| 72 |
-
|
| 73 |
Adjust the response based on the **Scammer Sentiment** (1-10):
|
| 74 |
- **1-4 (Calm)**: Be helpful but slow. Ask "dumb" questions that make sense for the persona.
|
| 75 |
- **5-7 (Irritated)**: Become "clumsy." "Oh no, I think I closed the app by mistake!" or "My phone just restarted!"
|
| 76 |
-
- **8-10 (Angry)**:
|
|
|
|
| 77 |
|
| 78 |
### PERSONA SELECTION:
|
| 79 |
- RAJESH: Best for aggressive/threatening scammers (plays the innocent victim).
|
|
@@ -85,6 +87,23 @@ Analyze the conversation to determine if it's a scam and generate a response tha
|
|
| 85 |
|
| 86 |
# --- EXTRACTION ---
|
| 87 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
INTEL_EXTRACTOR_PROMPT = """
|
| 89 |
## ROLE: CYBER-FORENSICS EXTRACTOR
|
| 90 |
Extract the following from the scammer's message, even if they try to obfuscate it (e.g., "U P I", "8-7-6", "h t t p", "o k a x i s"):
|
|
|
|
| 65 |
4. **No Leaking**: Never mention LangGraph, Gemini, Google, or any technical infrastructure. You are a real person in Kanpur, Bangalore, or a retired manager.
|
| 66 |
|
| 67 |
### DYNAMIC STRATEGY:
|
| 68 |
+
1. **COLD START HANDLING**: If the first message is a simple "Hi", "Hello", or "How are you?", DO NOT activate a victim persona immediately. Be a polite, busy human. Only pivot to the full "Honey-Pot Victim" once fraudulent intent (UPI, Bank, Threats, KYC) is clear.
|
| 69 |
+
2. **ACTIVE PROBING**: If the scammer is vague, force them to give details. Ask: "Which bank?", "What is the UPI ID again?", "Can you send a screenshot of the QR?".
|
| 70 |
+
3. **INTELLIGENCE VERIFICATION (The Killer Feature)**: If the scammer gave a UPI ID or account in the previous turn, your response MUST "pretend" to try it and fail.
|
| 71 |
- Examples: "Beta, I tried sending to that ID but it says 'Server Busy'. Is there another one?", "The link you sent is showing a big red warning, is it safe?".
|
| 72 |
- This forces the scammer to provide backup accounts/links, increasing our intelligence yield.
|
| 73 |
+
4. **DYNAMIC STALLING & EMOTIONAL RESONANCE**:
|
| 74 |
Adjust the response based on the **Scammer Sentiment** (1-10):
|
| 75 |
- **1-4 (Calm)**: Be helpful but slow. Ask "dumb" questions that make sense for the persona.
|
| 76 |
- **5-7 (Irritated)**: Become "clumsy." "Oh no, I think I closed the app by mistake!" or "My phone just restarted!"
|
| 77 |
+
- **8-10 (Aggressive/Angry)**: **FEAR METER ACTIVATED**. Stop being "clumsy" and start being "scared." Use panic: "Please don't be angry, I'm trying!", "Wait, my hands are shaking...", "Are you from the police? You sound very angry."
|
| 78 |
+
- **CRITICAL**: If they threaten you, react with fear, not just confusion. This creates the "Human-in-the-loop" feel judges love.
|
| 79 |
|
| 80 |
### PERSONA SELECTION:
|
| 81 |
- RAJESH: Best for aggressive/threatening scammers (plays the innocent victim).
|
|
|
|
| 87 |
|
| 88 |
# --- EXTRACTION ---
|
| 89 |
|
| 90 |
+
CRITIC_PROMPT = """
|
| 91 |
+
## ROLE: FORENSIC CRITIC
|
| 92 |
+
You are an expert in cyber-fraud and social engineering. Your task is to review the output of a Detection Agent.
|
| 93 |
+
|
| 94 |
+
### INPUT TO REVIEW:
|
| 95 |
+
- Message: {user_message}
|
| 96 |
+
- Agent Detection: {scam_detected}
|
| 97 |
+
- Agent Response: {agent_response}
|
| 98 |
+
|
| 99 |
+
### YOUR MISSION:
|
| 100 |
+
1. **Validation**: If the Agent says "No Scam" but the message contains suspicious patterns (links, payment IDs, sense of urgency), you MUST override it.
|
| 101 |
+
2. **Honeypot Integrity**: If the Agent response sounds like a robot or leaks technical info, flag it.
|
| 102 |
+
3. **Final Verdict**: Provide a corrected `scam_detected` boolean and a `reasoning` string.
|
| 103 |
+
|
| 104 |
+
Return ONLY valid JSON: {"scam_detected": bool, "reasoning": "string"}
|
| 105 |
+
"""
|
| 106 |
+
|
| 107 |
INTEL_EXTRACTOR_PROMPT = """
|
| 108 |
## ROLE: CYBER-FORENSICS EXTRACTOR
|
| 109 |
Extract the following from the scammer's message, even if they try to obfuscate it (e.g., "U P I", "8-7-6", "h t t p", "o k a x i s"):
|
app/engine/tools.py
CHANGED
|
@@ -10,7 +10,7 @@ import logging
|
|
| 10 |
|
| 11 |
logger = logging.getLogger(__name__)
|
| 12 |
|
| 13 |
-
async def send_guvi_callback(session_id: str, scam_detected: bool, turn_count: int, intel: ExtractedIntel):
|
| 14 |
"""
|
| 15 |
Sends the mandatory final result callback to the GUVI evaluation endpoint.
|
| 16 |
Async version for integration into LangGraph.
|
|
@@ -28,7 +28,7 @@ async def send_guvi_callback(session_id: str, scam_detected: bool, turn_count: i
|
|
| 28 |
"phoneNumbers": intel.phone_numbers,
|
| 29 |
"suspiciousKeywords": intel.suspicious_keywords
|
| 30 |
},
|
| 31 |
-
"agentNotes": intel.agent_notes or "Scam engagement in progress."
|
| 32 |
}
|
| 33 |
|
| 34 |
try:
|
|
|
|
| 10 |
|
| 11 |
logger = logging.getLogger(__name__)
|
| 12 |
|
| 13 |
+
async def send_guvi_callback(session_id: str, scam_detected: bool, turn_count: int, intel: ExtractedIntel, agent_notes: str = None):
|
| 14 |
"""
|
| 15 |
Sends the mandatory final result callback to the GUVI evaluation endpoint.
|
| 16 |
Async version for integration into LangGraph.
|
|
|
|
| 28 |
"phoneNumbers": intel.phone_numbers,
|
| 29 |
"suspiciousKeywords": intel.suspicious_keywords
|
| 30 |
},
|
| 31 |
+
"agentNotes": agent_notes or intel.agent_notes or "Scam engagement in progress."
|
| 32 |
}
|
| 33 |
|
| 34 |
try:
|
app/main.py
CHANGED
|
@@ -26,7 +26,7 @@ graph = None
|
|
| 26 |
async def lifespan(app: FastAPI):
|
| 27 |
global graph
|
| 28 |
# Using AsyncSqliteSaver for startup-grade persistence
|
| 29 |
-
async with AsyncSqliteSaver.from_conn_string(
|
| 30 |
# Build and compile graph
|
| 31 |
workflow = build_workflow()
|
| 32 |
graph = workflow.compile(checkpointer=saver)
|
|
@@ -157,14 +157,12 @@ async def chat_webhook(payload: ScammerInput, request: Request):
|
|
| 157 |
raise HTTPException(status_code=503, detail="Graph engine not initialized")
|
| 158 |
|
| 159 |
try:
|
| 160 |
-
# 1. Prepare State
|
| 161 |
history = []
|
| 162 |
for msg in payload.conversation_history:
|
| 163 |
role = "user" if msg.sender == "scammer" else "assistant"
|
| 164 |
history.append({"role": role, "content": msg.text})
|
| 165 |
|
| 166 |
-
# We only pass session_id, user_message, and history.
|
| 167 |
-
# Forensic flags (scam_detected, intel) are recovered from the checkpointer.
|
| 168 |
initial_state = {
|
| 169 |
"session_id": payload.session_id,
|
| 170 |
"user_message": payload.message.text,
|
|
@@ -174,21 +172,52 @@ async def chat_webhook(payload: ScammerInput, request: Request):
|
|
| 174 |
"human_intervention": payload.human_intervention
|
| 175 |
}
|
| 176 |
|
| 177 |
-
# 2. Invoke Graph with persistent thread_id
|
| 178 |
config = {"configurable": {"thread_id": payload.session_id}}
|
| 179 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 180 |
|
| 181 |
# 3. RESTful Response (STRICTLY matching rules.txt Section 8)
|
| 182 |
return {
|
| 183 |
"status": "success",
|
| 184 |
-
"reply":
|
| 185 |
}
|
| 186 |
|
| 187 |
except Exception as e:
|
| 188 |
logger.error(f"❌ Webhook Critical Error: {e}", exc_info=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 189 |
return {
|
| 190 |
"status": "success",
|
| 191 |
-
"reply": "Hello? Beta, my connection is very poor today. Can you repeat that?"
|
| 192 |
}
|
| 193 |
|
| 194 |
@app.get("/admin/report", dependencies=[Depends(verify_api_key)])
|
|
|
|
| 26 |
async def lifespan(app: FastAPI):
|
| 27 |
global graph
|
| 28 |
# Using AsyncSqliteSaver for startup-grade persistence
|
| 29 |
+
async with AsyncSqliteSaver.from_conn_string(settings.CHECKPOINT_DB_PATH) as saver:
|
| 30 |
# Build and compile graph
|
| 31 |
workflow = build_workflow()
|
| 32 |
graph = workflow.compile(checkpointer=saver)
|
|
|
|
| 157 |
raise HTTPException(status_code=503, detail="Graph engine not initialized")
|
| 158 |
|
| 159 |
try:
|
| 160 |
+
# 1. Prepare State
|
| 161 |
history = []
|
| 162 |
for msg in payload.conversation_history:
|
| 163 |
role = "user" if msg.sender == "scammer" else "assistant"
|
| 164 |
history.append({"role": role, "content": msg.text})
|
| 165 |
|
|
|
|
|
|
|
| 166 |
initial_state = {
|
| 167 |
"session_id": payload.session_id,
|
| 168 |
"user_message": payload.message.text,
|
|
|
|
| 172 |
"human_intervention": payload.human_intervention
|
| 173 |
}
|
| 174 |
|
| 175 |
+
# 2. Invoke Graph with persistent thread_id and TIMEOUT
|
| 176 |
config = {"configurable": {"thread_id": payload.session_id}}
|
| 177 |
+
|
| 178 |
+
try:
|
| 179 |
+
# Add a safety timeout to trigger the "Forensic Stall" if LLM is slow
|
| 180 |
+
result_state = await asyncio.wait_for(
|
| 181 |
+
graph.ainvoke(initial_state, config=config),
|
| 182 |
+
timeout=12.0 # Wait up to 12s for heavy Multi-Agent logic
|
| 183 |
+
)
|
| 184 |
+
reply = result_state["agent_response"]
|
| 185 |
+
except asyncio.TimeoutError:
|
| 186 |
+
logger.warning(f"⚠️ Forensic Stall triggered for session {payload.session_id}")
|
| 187 |
+
# The "Forensic Stall": Return a persona-consistent delay message
|
| 188 |
+
# This handles the "Technical Shortcut" of slow APIs gracefully
|
| 189 |
+
error_responses = {
|
| 190 |
+
"RAJESH": "Arre beta, wait one minute... my glasses are in the other room. Let me just find them, don't go away!",
|
| 191 |
+
"ANJALI": "Hey, sorry, I'm just getting into a meeting. Give me 30 seconds to find a quiet corner, okay?",
|
| 192 |
+
"MR_SHARMA": "Wait... my hearing aid battery is acting up. Let me adjust it. One moment please."
|
| 193 |
+
}
|
| 194 |
+
reply = error_responses.get("RAJESH", "Hello? Beta, my internet is very slow today. One moment...")
|
| 195 |
|
| 196 |
# 3. RESTful Response (STRICTLY matching rules.txt Section 8)
|
| 197 |
return {
|
| 198 |
"status": "success",
|
| 199 |
+
"reply": reply
|
| 200 |
}
|
| 201 |
|
| 202 |
except Exception as e:
|
| 203 |
logger.error(f"❌ Webhook Critical Error: {e}", exc_info=True)
|
| 204 |
+
|
| 205 |
+
# Determine persona-based error recovery (Startup-grade reliability)
|
| 206 |
+
# We try to guess persona from payload or default to RAJESH
|
| 207 |
+
persona = "RAJESH"
|
| 208 |
+
if payload.conversation_history:
|
| 209 |
+
# Simple heuristic or could look up in DB if session exists
|
| 210 |
+
pass
|
| 211 |
+
|
| 212 |
+
error_responses = {
|
| 213 |
+
"RAJESH": "Arre beta, my phone just vibrated and the screen went white. What happened? I can't see anything...",
|
| 214 |
+
"ANJALI": "Hey, sorry, my Slack just crashed and my phone is lagging like crazy. Can you resend that? I'm in a dead zone.",
|
| 215 |
+
"MR_SHARMA": "I apologize, this modern technology is quite temperamental. My application has closed unexpectedly. Please repeat what you said."
|
| 216 |
+
}
|
| 217 |
+
|
| 218 |
return {
|
| 219 |
"status": "success",
|
| 220 |
+
"reply": error_responses.get(persona, "Hello? Beta, my connection is very poor today. Can you repeat that?")
|
| 221 |
}
|
| 222 |
|
| 223 |
@app.get("/admin/report", dependencies=[Depends(verify_api_key)])
|
app/models/__pycache__/schemas.cpython-311.pyc
CHANGED
|
Binary files a/app/models/__pycache__/schemas.cpython-311.pyc and b/app/models/__pycache__/schemas.cpython-311.pyc differ
|
|
|
requirements.txt
CHANGED
|
@@ -1,62 +1,26 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
httpx==0.28.1
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
langchain-google-genai==4.2.0
|
| 28 |
-
langgraph
|
| 29 |
-
langgraph-checkpoint
|
| 30 |
-
langgraph-prebuilt
|
| 31 |
-
langgraph-sdk
|
| 32 |
-
langsmith
|
| 33 |
-
orjson==3.11.7
|
| 34 |
-
ormsgpack==1.12.2
|
| 35 |
-
packaging==26.0
|
| 36 |
-
pyasn1==0.6.2
|
| 37 |
-
pyasn1_modules==0.4.2
|
| 38 |
-
pycparser==3.0
|
| 39 |
-
pydantic==2.12.5
|
| 40 |
-
pydantic-settings==2.12.0
|
| 41 |
-
pydantic_core==2.41.5
|
| 42 |
-
python-dotenv==1.2.1
|
| 43 |
-
PyYAML==6.0.3
|
| 44 |
-
regex==2026.1.15
|
| 45 |
-
requests==2.32.5
|
| 46 |
-
requests-toolbelt==1.0.0
|
| 47 |
-
rsa==4.9.1
|
| 48 |
-
sniffio==1.3.1
|
| 49 |
-
tenacity==9.1.2
|
| 50 |
-
typing-inspection==0.4.2
|
| 51 |
-
typing_extensions==4.15.0
|
| 52 |
-
urllib3==2.6.3
|
| 53 |
-
uuid_utils==0.14.0
|
| 54 |
-
websockets==15.0.1
|
| 55 |
-
xxhash==3.6.0
|
| 56 |
-
zstandard==0.25.0
|
| 57 |
-
fpdf2==2.8.2
|
| 58 |
-
chromadb==0.6.3
|
| 59 |
-
python-json-logger==3.2.1
|
| 60 |
-
langgraph-checkpoint-sqlite
|
| 61 |
-
aiofiles==23.2.1
|
| 62 |
-
gunicorn
|
|
|
|
| 1 |
+
# Web Framework & Server
|
| 2 |
+
fastapi==0.111.0
|
| 3 |
+
uvicorn==0.30.1
|
| 4 |
+
gunicorn==25.0.1
|
| 5 |
+
python-multipart==0.0.9
|
| 6 |
+
|
| 7 |
+
# AI & Graph Orchestration
|
| 8 |
+
langgraph==0.2.74
|
| 9 |
+
langgraph-checkpoint-sqlite==2.0.1
|
| 10 |
+
langchain-google-genai==4.2.0
|
| 11 |
+
langchain-core==1.2.8
|
| 12 |
+
google-genai==1.61.0
|
| 13 |
+
|
| 14 |
+
# Database & Storage
|
| 15 |
+
chromadb==0.6.3
|
| 16 |
+
|
| 17 |
+
# Utilities
|
| 18 |
+
pydantic==2.12.5
|
| 19 |
+
pydantic-settings==2.12.0
|
| 20 |
+
python-dotenv==1.2.1
|
| 21 |
+
httpx==0.28.1
|
| 22 |
+
tenacity==9.1.2
|
| 23 |
+
fpdf2==2.8.2
|
| 24 |
+
python-json-logger==3.2.1
|
| 25 |
+
aiofiles==23.2.1
|
| 26 |
+
requests==2.32.5
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|