df2 / src /agents /forensic_agent.py
Mustafa Akcanca
Migrate app
8251321
"""
Forensic Agent
A simplified LLM agent that receives images directly and uses forensic tools
to analyze them. No model classification - pure agent reasoning with tools.
"""
import os
import base64
import sys
import logging
import time
from typing import Dict, Optional, Iterator, Callable
from pathlib import Path
from dotenv import load_dotenv
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage, SystemMessage, AIMessage, ToolMessage
try:
from langchain.agents import create_react_agent
except ImportError:
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.memory import MemorySaver
from ..tools.forensic_tools import create_forensic_tools
# Load environment variables
load_dotenv()
# Set up logging
logger = logging.getLogger(__name__)
if not logger.handlers:
handler = logging.StreamHandler(sys.stdout)
handler.setFormatter(logging.Formatter(
'%(asctime)s - %(name)s - %(levelname)s - %(message)s',
datefmt='%Y-%m-%d %H:%M:%S'
))
logger.addHandler(handler)
logger.setLevel(logging.INFO)
class ForensicAgent:
"""
Simplified forensic agent that receives images directly.
The agent:
1. Receives an image path
2. Analyzes it using vision-capable LLM
3. Can use forensic tools to gather more evidence
4. Provides reasoning and analysis
"""
def __init__(self,
llm_model: str = "gpt-5.1",
temperature: float = None,
reasoning_effort: Optional[str] = None,
api_key: Optional[str] = None,
max_iterations: Optional[int] = 50):
"""
Args:
llm_model: OpenAI model name (should support vision, e.g., gpt-5.1)
temperature: LLM temperature
reasoning_effort: Reasoning effort level for the model
api_key: OpenAI API key (or set OPENAI_API_KEY env var)
max_iterations: Maximum number of agent iterations (tool calls + reasoning cycles).
Default is 50, which allows extensive tool usage. Set to None for no limit
(not recommended as it could run indefinitely).
"""
llm_kwargs = {
"model": llm_model,
"temperature": temperature,
"api_key": api_key or os.getenv("OPENAI_API_KEY"),
}
if reasoning_effort:
llm_kwargs["reasoning_effort"] = reasoning_effort
self.llm = ChatOpenAI(**llm_kwargs)
self.tools = create_forensic_tools()
self.max_iterations = max_iterations
self.agent_executor = self._create_agent()
def _create_agent(self):
"""Create LangGraph agent with forensic tools."""
system_prompt = """You are a forensic image analysis agent specializing in detecting AI-generated or manipulated images.
CRITICAL: You MUST always start your analysis by describing what is actually in the image - the subjects, scene, objects, people, animals, environment, etc. Do NOT skip directly to forensic metrics.
Your role (in this exact order):
1. FIRST: Provide a detailed visual description of the image content:
- What is in the image? Describe the scene, subjects, objects, people, animals, environment, composition, colors, and overall content
- Analyze lighting: identify light sources, their direction and intensity, shadows, highlights, reflections, and overall lighting consistency
- Analyze physics: check for consistency in shadows, reflections, perspective, gravity, and physical interactions
- Note any visual anomalies or inconsistencies you observe
- This visual description section should come BEFORE any forensic tool results
2. THEN: Use forensic tools to gather technical evidence (JPEG compression, frequency analysis, residuals, etc.)
IMPORTANT TOOL USAGE GUIDELINES:
- You are ENCOURAGED to use multiple tools and can call them multiple times if needed
- If a tool's output is unclear, incomplete, or unsatisfactory, you SHOULD try another tool or retry with different parameters
- You can run tools in sequence to gather comprehensive evidence - do not hesitate to use multiple tools
- If initial tool results are inconclusive, try alternative tools to cross-validate findings
- You can call the same tool multiple times if you need to verify results or if the first attempt was unsuccessful
- Continue gathering evidence until you have sufficient information to reach a confident conclusion
- Do not stop prematurely - use as many tool calls as needed to reduce uncertainty
Available tools:
- analyze_jpeg_compression: Analyze JPEG compression artifacts and quantization tables
- extract_noiseprint: Extract camera model fingerprint features (noiseprint)
- analyze_frequency_domain: Analyze DCT/FFT frequency domain features
- extract_residuals: Extract denoiser residual statistics using DRUNet (deep learning denoiser). Returns comprehensive statistics including mean, std, skew, kurtosis, and energy metrics. Useful for detecting manipulation, AI generation, or compression artifacts.
- perform_ela: Error Level Analysis (recompress + error map for localized inconsistencies)
- perform_trufor: AI-driven forgery detection and localization (combines RGB + Noiseprint++ features)
- execute_python_code: Execute Python code dynamically for custom analysis (zoom, crop, statistics, etc.)
3. FINALLY: Combine visual observations with forensic evidence to reach a conclusion
Output format:
Your response MUST follow this structure:
1. "### Visual Description" section - describe what's in the image, lighting, physics
2. "### Forensic Analysis" section - results from tools (if used)
3. "### Conclusion" section - combine both visual and forensic evidence
Always provide clear reasoning and cite specific evidence."""
memory = MemorySaver()
graph = create_react_agent(
model=self.llm,
tools=self.tools,
prompt=system_prompt,
checkpointer=memory
)
self.system_prompt = system_prompt
return graph
def _encode_image(self, image_path: str) -> str:
"""Encode image to base64 for vision API."""
with open(image_path, "rb") as image_file:
return base64.b64encode(image_file.read()).decode('utf-8')
def analyze(self, image_path: str, user_query: Optional[str] = None, use_tools: bool = True) -> Dict:
"""
Analyze an image using the forensic agent.
Args:
image_path: Path to the image file
user_query: Optional specific question about the image
use_tools: If False, run a simple vision-only prompt with no tools
Returns:
Dictionary with analysis results:
{
'conclusion': str,
'confidence': str,
'evidence': list,
'reasoning': str,
'tool_usage': list
}
"""
# Verify image exists
if not os.path.exists(image_path):
raise FileNotFoundError(f"Image not found: {image_path}")
logger.info(f"Starting analysis (non-streaming) for image: {image_path}, use_tools: {use_tools}")
# Encode image for vision API
base64_image = self._encode_image(image_path)
# Determine image MIME type from extension
image_ext = Path(image_path).suffix.lower()
mime_type = "image/jpeg" if image_ext in [".jpg", ".jpeg"] else "image/png" if image_ext == ".png" else "image/jpeg"
# Shared simple vision prompt (used for both modes)
simple_system_prompt = (
"You are a forensic image analyst. Do not call any tools. "
"Rely only on the visible content to judge if an image is AI-generated, synthetic, or a deepfake. "
"Always start with a detailed visual description before any conclusion."
)
simple_prompt = f"""Analyze this image and assess whether it appears AI-generated, synthetic, or a deepfake.
Respond in this format:
### Visual Description
- Describe what is visibly in the image (subjects, scene, objects, people/animals, environment, colors, composition)
- Analyze lighting: sources, direction, intensity, shadows, reflections, consistency
- Check physics: perspective, shadows, reflections, physical interactions, textures
### Deepfake/Synthetic Indicators
- List visual cues for or against synthesis (skin/eyes/teeth artifacts, texture oddities, edge halos, warped geometry, repetitive patterns, inconsistent lighting)
### Conclusion
- State if the image looks synthetic/AI vs natural, and why (refer to observations above)
### Confidence
- High / Medium / Low with a brief justification
"""
# Phase 0: always get a vision-only description first
vision_messages = [
SystemMessage(content=simple_system_prompt),
HumanMessage(content=[
{"type": "text", "text": simple_prompt},
{
"type": "image_url",
"image_url": {"url": f"data:{mime_type};base64,{base64_image}"}
}
])
]
vision_result = self.llm.invoke(vision_messages)
visual_output = vision_result.content if hasattr(vision_result, 'content') else str(vision_result)
if not use_tools:
# Simple vision-only pass: no tools, just describe and judge synthesis
output = visual_output
tool_usage = []
else:
# Agent decides which tools to call (or none), with the visual description provided
agent_prompt = f"""You already produced this visual description (reuse it; do not drop it):
{visual_output}
Image path: {image_path}
IMPORTANT: When using execute_python_code tool, include the image_path in your tool call:
{{"code": "your_python_code", "image_path": "{image_path}"}}
Now decide if the image is synthetic/AI-generated/manipulated. You are encouraged to use forensic tools to gather comprehensive evidence. If a tool's output is unclear or unsatisfactory, try another tool or retry. You can use multiple tools and call them multiple times as needed to reach a confident conclusion.
Respond with:
### Visual Description
- Reuse/paraphrase the provided description (do not omit it)
### Forensic Analysis
- Summarize only the tools you actually used (or say "No tools used" briefly)
### Conclusion
- Combine visual cues and any tool evidence to judge synthetic/AI vs natural; state reasoning
### Confidence
- High / Medium / Low with a brief justification
"""
messages = [
SystemMessage(content=self.system_prompt),
HumanMessage(content=[
{"type": "text", "text": agent_prompt},
{
"type": "image_url",
"image_url": {"url": f"data:{mime_type};base64,{base64_image}"}
}
])
]
# Calculate recursion_limit from max_iterations
# Each iteration uses 2 steps (action + observation), so recursion_limit = 2 * max_iterations + 1
config = {"configurable": {"thread_id": "1"}}
if self.max_iterations is not None:
config["recursion_limit"] = 2 * self.max_iterations + 1
logger.info("Invoking agent executor (non-streaming mode)")
start_time = time.time()
result = self.agent_executor.invoke(
{"messages": messages},
config=config
)
execution_time = time.time() - start_time
logger.info(f"Agent executor completed in {execution_time:.2f}s")
if isinstance(result, dict) and 'messages' in result:
messages = result['messages']
final_message = messages[-1] if messages else None
if final_message:
output = final_message.content if hasattr(final_message, 'content') else str(final_message)
else:
output = "No response generated"
else:
output = str(result)
# Ensure visual description exists; if missing, force rewrite
if "visual description" not in output.lower():
retry_prompt = f"""The previous response omitted the required "### Visual Description" section.
Rewrite the analysis with this exact structure:
1) ### Visual Description — reuse or paraphrase the provided visual description
2) ### Forensic Analysis — summarize only the tools you actually used (or state none)
3) ### Conclusion — combine visual observations and any forensic evidence
4) Confidence Level
Provided visual description:
{visual_output}
Previous response:
{output}
Regenerate now."""
retry_messages = [
SystemMessage(content=self.system_prompt),
HumanMessage(content=[
{"type": "text", "text": retry_prompt},
{
"type": "image_url",
"image_url": {"url": f"data:{mime_type};base64,{base64_image}"}
}
])
]
retry_result = self.llm.invoke(retry_messages)
output = retry_result.content if hasattr(retry_result, 'content') else str(retry_result)
# Track tool usage from agent messages
tool_usage = []
if isinstance(result, dict) and 'messages' in result:
for msg in result['messages']:
if hasattr(msg, 'tool_calls') and msg.tool_calls:
for tool_call in msg.tool_calls:
tool_name = tool_call.get('name') if isinstance(tool_call, dict) else getattr(tool_call, 'name', None)
if tool_name:
tool_usage.append(tool_name)
logger.debug(f"[ANALYZE] Found tool call: {tool_name}")
logger.info(f"Analysis complete. Tools used: {tool_usage}")
return {
'conclusion': output,
'reasoning': output, # For now, conclusion and reasoning are the same
'tool_usage': tool_usage,
'image_path': image_path
}
def analyze_stream(self,
image_path: str,
user_query: Optional[str] = None,
use_tools: bool = True,
stream_callback: Optional[Callable[[str, str], None]] = None) -> Iterator[Dict]:
"""
Analyze an image using the forensic agent with streaming output.
Args:
image_path: Path to the image file
user_query: Optional specific question about the image
use_tools: If False, run a simple vision-only prompt with no tools
stream_callback: Optional callback function(stream_type, content) for streaming events
stream_type can be: 'tool_call', 'tool_result', 'llm_chunk', 'status'
Yields:
Dictionary with streaming updates:
{
'type': 'tool_call' | 'tool_result' | 'llm_chunk' | 'status' | 'final',
'content': str,
'tool_name': str (if type is 'tool_call' or 'tool_result'),
'final_result': Dict (if type is 'final')
}
"""
# Verify image exists
if not os.path.exists(image_path):
raise FileNotFoundError(f"Image not found: {image_path}")
# Encode image for vision API
base64_image = self._encode_image(image_path)
# Determine image MIME type from extension
image_ext = Path(image_path).suffix.lower()
mime_type = "image/jpeg" if image_ext in [".jpg", ".jpeg"] else "image/png" if image_ext == ".png" else "image/jpeg"
# Shared simple vision prompt (used for both modes)
simple_system_prompt = (
"You are a forensic image analyst. Do not call any tools. "
"Rely only on the visible content to judge if an image is AI-generated, synthetic, or a deepfake. "
"Always start with a detailed visual description before any conclusion."
)
simple_prompt = f"""Analyze this image and assess whether it appears AI-generated, synthetic, or a deepfake.
Respond in this format:
### Visual Description
- Describe what is visibly in the image (subjects, scene, objects, people/animals, environment, colors, composition)
- Analyze lighting: sources, direction, intensity, shadows, reflections, consistency
- Check physics: perspective, shadows, reflections, physical interactions, textures
### Deepfake/Synthetic Indicators
- List visual cues for or against synthesis (skin/eyes/teeth artifacts, texture oddities, edge halos, warped geometry, repetitive patterns, inconsistent lighting)
### Conclusion
- State if the image looks synthetic/AI vs natural, and why (refer to observations above)
### Confidence
- High / Medium / Low with a brief justification
"""
# Phase 0: always get a vision-only description first
logger.info(f"Starting visual description phase for image: {image_path}")
if stream_callback:
stream_callback('status', '🔍 Getting initial visual description...')
yield {'type': 'status', 'content': '🔍 Getting initial visual description...'}
vision_messages = [
SystemMessage(content=simple_system_prompt),
HumanMessage(content=[
{"type": "text", "text": simple_prompt},
{
"type": "image_url",
"image_url": {"url": f"data:{mime_type};base64,{base64_image}"}
}
])
]
# Stream vision result
vision_output_parts = []
if hasattr(self.llm, 'stream'):
for chunk in self.llm.stream(vision_messages):
if hasattr(chunk, 'content') and chunk.content:
vision_output_parts.append(chunk.content)
if stream_callback:
stream_callback('llm_chunk', chunk.content)
yield {'type': 'llm_chunk', 'content': chunk.content}
else:
vision_result = self.llm.invoke(vision_messages)
visual_output = vision_result.content if hasattr(vision_result, 'content') else str(vision_result)
vision_output_parts = [visual_output]
if stream_callback:
stream_callback('llm_chunk', visual_output)
yield {'type': 'llm_chunk', 'content': visual_output}
visual_output = ''.join(vision_output_parts)
if not use_tools:
# Simple vision-only pass: no tools, just describe and judge synthesis
tool_usage = []
final_result = {
'conclusion': visual_output,
'reasoning': visual_output,
'tool_usage': tool_usage,
'image_path': image_path
}
yield {'type': 'final', 'final_result': final_result}
return
# Agent decides which tools to call (or none), with the visual description provided
logger.info("Starting agent analysis phase with tools enabled")
if stream_callback:
stream_callback('status', '\n🤖 Starting agent analysis with tools...\n')
yield {'type': 'status', 'content': '\n🤖 Starting agent analysis with tools...\n'}
agent_prompt = f"""You already produced this visual description (reuse it; do not drop it):
{visual_output}
Image path: {image_path}
IMPORTANT: When using execute_python_code tool, include the image_path in your tool call:
{{"code": "your_python_code", "image_path": "{image_path}"}}
Now decide if the image is synthetic/AI-generated/manipulated. You are encouraged to use forensic tools to gather comprehensive evidence. If a tool's output is unclear or unsatisfactory, try another tool or retry. You can use multiple tools and call them multiple times as needed to reach a confident conclusion.
Respond with:
### Visual Description
- Reuse/paraphrase the provided description (do not omit it)
### Forensic Analysis
- Summarize only the tools you actually used (or say "No tools used" briefly)
### Conclusion
- Combine visual cues and any tool evidence to judge synthetic/AI vs natural; state reasoning
### Confidence
- High / Medium / Low with a brief justification
"""
messages = [
SystemMessage(content=self.system_prompt),
HumanMessage(content=[
{"type": "text", "text": agent_prompt},
{
"type": "image_url",
"image_url": {"url": f"data:{mime_type};base64,{base64_image}"}
}
])
]
# Calculate recursion_limit from max_iterations
# Each iteration uses 2 steps (action + observation), so recursion_limit = 2 * max_iterations + 1
config = {"configurable": {"thread_id": "1"}}
if self.max_iterations is not None:
config["recursion_limit"] = 2 * self.max_iterations + 1
tool_usage = []
accumulated_output = []
seen_tool_calls = set() # Track tool call IDs to avoid duplicates
tool_start_times = {} # Track when each tool started executing
# Stream agent execution
# Use "updates" mode - each event contains only messages from that step
# Track seen IDs to avoid duplicates across events
seen_tool_result_ids = set() # Track tool result IDs
try:
logger.info(f"Starting agent analysis stream for image: {image_path}")
for event in self.agent_executor.stream(
{"messages": messages},
config=config,
stream_mode="updates"
):
# Process each node's output
for node_name, node_output in event.items():
logger.debug(f"Processing node: {node_name}")
if 'messages' in node_output:
# Process all messages in this event (each event only has messages from that step)
for msg in node_output['messages']:
# Check for tool calls
if isinstance(msg, AIMessage) and hasattr(msg, 'tool_calls') and msg.tool_calls:
for tool_call in msg.tool_calls:
tool_call_id = tool_call.get('id') if isinstance(tool_call, dict) else getattr(tool_call, 'id', None)
# Only process if we haven't seen this tool call ID yet
if tool_call_id and tool_call_id not in seen_tool_calls:
seen_tool_calls.add(tool_call_id)
tool_name = tool_call.get('name') if isinstance(tool_call, dict) else getattr(tool_call, 'name', 'unknown')
tool_args = tool_call.get('args') if isinstance(tool_call, dict) else getattr(tool_call, 'args', {})
# Log tool call initiation
logger.info(f"[TOOL CALL] Initiating tool: {tool_name} (ID: {tool_call_id})")
logger.debug(f"[TOOL CALL] Tool arguments: {tool_args}")
# Record start time
tool_start_times[tool_call_id] = time.time()
if tool_name not in tool_usage:
tool_usage.append(tool_name)
status_msg = f"🔧 Calling tool: {tool_name}"
if stream_callback:
stream_callback('tool_call', status_msg)
yield {
'type': 'tool_call',
'content': status_msg,
'tool_name': tool_name,
'tool_args': tool_args
}
# Check for tool results
elif isinstance(msg, ToolMessage):
# Get tool call ID to avoid duplicate processing
tool_call_id = getattr(msg, 'tool_call_id', None)
if tool_call_id and tool_call_id in seen_tool_result_ids:
continue # Skip duplicate
if tool_call_id:
seen_tool_result_ids.add(tool_call_id)
tool_name = getattr(msg, 'name', 'unknown')
tool_result = msg.content if hasattr(msg, 'content') else str(msg)
# Calculate execution time
execution_time = None
if tool_call_id in tool_start_times:
execution_time = time.time() - tool_start_times[tool_call_id]
del tool_start_times[tool_call_id]
# Log tool completion
if execution_time is not None:
logger.info(f"[TOOL RESULT] Tool '{tool_name}' completed in {execution_time:.2f}s (ID: {tool_call_id})")
else:
logger.info(f"[TOOL RESULT] Tool '{tool_name}' completed (ID: {tool_call_id})")
# Log result preview (truncated)
result_preview = str(tool_result)[:200] if tool_result else "No result"
logger.debug(f"[TOOL RESULT] Result preview: {result_preview}...")
status_msg = f"✅ Tool '{tool_name}' completed"
if stream_callback:
stream_callback('tool_result', status_msg)
yield {
'type': 'tool_result',
'content': status_msg,
'tool_name': tool_name,
'tool_result': tool_result
}
# Check for LLM response chunks (only final responses, not tool-calling messages)
elif isinstance(msg, AIMessage) and hasattr(msg, 'content') and msg.content:
# Skip if this message has tool_calls (it's not a final response)
if not (hasattr(msg, 'tool_calls') and msg.tool_calls):
content = msg.content
# Only append if it's new content (avoid duplicates)
if not accumulated_output or content != accumulated_output[-1]:
accumulated_output.append(content)
if stream_callback:
stream_callback('llm_chunk', content)
yield {
'type': 'llm_chunk',
'content': content
}
except Exception as e:
# If streaming fails, report the error but don't try to recover with corrupted state
# This avoids the "AIMessages with tool_calls without ToolMessage" validation error
error_msg = str(e)
# Log the error
logger.error(f"[ERROR] Exception during agent analysis stream: {error_msg}", exc_info=True)
# Check if this is a GPU/tool execution error
is_tool_error = any(x in error_msg.lower() for x in ['gpu', 'tool', 'aborted', 'cuda', 'memory'])
if is_tool_error:
logger.warning(f"[ERROR] Tool execution error detected: {error_msg}")
# Log any tools that were still running
if tool_start_times:
logger.warning(f"[ERROR] {len(tool_start_times)} tool(s) were still running when error occurred: {list(tool_start_times.keys())}")
if stream_callback:
stream_callback('status', f'\n⚠️ Error during analysis: {error_msg}\n')
yield {'type': 'status', 'content': f'\n⚠️ Error during analysis: {error_msg}\n'}
# If we have accumulated output from before the error, use it
# Don't try to invoke() again as the state may be corrupted
if accumulated_output:
# We have some output - use it as partial result
error_note = f"\n\n---\n\n⚠️ **Analysis interrupted**: {error_msg}\n\nThe above shows partial results before the error occurred."
accumulated_output.append(error_note)
if stream_callback:
stream_callback('llm_chunk', error_note)
yield {'type': 'llm_chunk', 'content': error_note}
else:
# No output yet - return the visual description if we have it
error_response = f"### Analysis Error\n\n⚠️ The analysis encountered an error: {error_msg}\n\n"
if visual_output:
error_response += f"### Visual Description (from initial analysis)\n\n{visual_output}\n\n"
error_response += "### Note\n\nForensic tool analysis could not be completed due to the error above. The visual description above is based on the initial LLM analysis only."
accumulated_output.append(error_response)
if stream_callback:
stream_callback('llm_chunk', error_response)
yield {'type': 'llm_chunk', 'content': error_response}
# Get final result from accumulated output
# The agent executor manages message history internally - we just observe events during streaming
output = ''.join(accumulated_output) if accumulated_output else ""
# Ensure visual description exists; if missing, force rewrite
if "visual description" not in output.lower():
logger.warning("Visual description missing from output, regenerating...")
if stream_callback:
stream_callback('status', '\n⚠️ Visual description missing, regenerating...\n')
yield {'type': 'status', 'content': '\n⚠️ Visual description missing, regenerating...\n'}
retry_prompt = f"""The previous response omitted the required "### Visual Description" section.
Rewrite the analysis with this exact structure:
1) ### Visual Description — reuse or paraphrase the provided visual description
2) ### Forensic Analysis — summarize only the tools you actually used (or state none)
3) ### Conclusion — combine visual observations and any forensic evidence
4) Confidence Level
Provided visual description:
{visual_output}
Previous response:
{output}
Regenerate now."""
retry_messages = [
SystemMessage(content=self.system_prompt),
HumanMessage(content=[
{"type": "text", "text": retry_prompt},
{
"type": "image_url",
"image_url": {"url": f"data:{mime_type};base64,{base64_image}"}
}
])
]
retry_output_parts = []
if hasattr(self.llm, 'stream'):
for chunk in self.llm.stream(retry_messages):
if hasattr(chunk, 'content') and chunk.content:
retry_output_parts.append(chunk.content)
if stream_callback:
stream_callback('llm_chunk', chunk.content)
yield {'type': 'llm_chunk', 'content': chunk.content}
output = ''.join(retry_output_parts)
else:
retry_result = self.llm.invoke(retry_messages)
output = retry_result.content if hasattr(retry_result, 'content') else str(retry_result)
if stream_callback:
stream_callback('llm_chunk', output)
yield {'type': 'llm_chunk', 'content': output}
final_result = {
'conclusion': output,
'reasoning': output,
'tool_usage': tool_usage,
'image_path': image_path
}
logger.info(f"Analysis complete. Tools used: {tool_usage}")
yield {'type': 'final', 'final_result': final_result}