EgoBench / run /utils.py
emodiary's picture
Initial complete EgoBench release
3c4b9d3 verified
Raw History Blame Contribute Delete
19.9 kB
import os
import sys
import json
import base64
import re
import time
import random
import mimetypes
from urllib.parse import urlparse
import requests
# Key: Add project root to Python module search path
current_file_path = os.path.abspath(__file__)
run_dir = os.path.dirname(current_file_path)
project_root = os.path.dirname(run_dir)
sys.path.insert(0, os.path.abspath(project_root))
from run.prompts import (
USER_CONTRADICTION_CHECK_PROMPT,
USER_RESPONSE_CORRECTION_PROMPT,
USER_TEXT_ONLY_PROMPT_EASY,
USER_TEXT_ONLY_PROMPT_HARD,
)
from run.apis import call_llm
# --- Media processing utility functions ---
def is_url(path):
"""Check if path is a URL"""
try:
result = urlparse(path)
return all([result.scheme, result.netloc])
except Exception:
return False
def get_media_type_from_url(url):
"""Get media type from URL"""
mime_type, _ = mimetypes.guess_type(url)
if mime_type and mime_type.startswith('video'):
return 'video'
elif mime_type and mime_type.startswith('image'):
return 'image'
else:
parsed_url = urlparse(url)
ext = os.path.splitext(parsed_url.path)[1].lower()
video_exts = ['.mp4', '.avi', '.mov', '.mkv', '.flv', '.wmv', '.webm']
if ext in video_exts:
return 'video'
return 'image'
def get_media_mime_type_from_url(url):
"""Get specific MIME type from URL"""
mime_type, _ = mimetypes.guess_type(url)
if mime_type:
return mime_type
parsed_url = urlparse(url)
ext = os.path.splitext(parsed_url.path)[1].lower()
mime_map = {
'.mp4': 'video/mp4',
'.avi': 'video/avi',
'.mov': 'video/quicktime',
'.mkv': 'video/x-matroska',
'.webm': 'video/webm',
'.jpg': 'image/jpeg',
'.jpeg': 'image/jpeg',
'.png': 'image/png',
'.gif': 'image/gif',
'.webp': 'image/webp'
}
return mime_map.get(ext, 'image/jpeg')
def encode_image(image_path):
"""Encode image file"""
if not image_path or not os.path.exists(image_path):
return None
with open(image_path, "rb") as f:
return base64.b64encode(f.read()).decode('utf-8')
def encode_video(video_path):
"""Encode video file"""
if not video_path or not os.path.exists(video_path):
return None
with open(video_path, "rb") as f:
return base64.b64encode(f.read()).decode('utf-8')
def download_media_from_url(url):
"""Download media content from URL"""
try:
response = requests.get(url)
response.raise_for_status()
return response.content
except Exception as e:
print(f"Failed to download media: {e}")
return None
def get_media_type(file_path):
"""Determine media type based on file extension"""
if is_url(file_path):
return get_media_type_from_url(file_path)
else:
mime_type, _ = mimetypes.guess_type(file_path)
if mime_type and mime_type.startswith('video'):
return 'video'
elif mime_type and mime_type.startswith('image'):
return 'image'
else:
ext = os.path.splitext(file_path)[1].lower()
video_exts = ['.mp4', '.avi', '.mov', '.mkv', '.flv', '.wmv', '.webm']
if ext in video_exts:
return 'video'
return 'image'
def get_media_mime_type(file_path):
"""Get specific MIME type"""
if is_url(file_path):
return get_media_mime_type_from_url(file_path)
else:
mime_type, _ = mimetypes.guess_type(file_path)
if mime_type:
return mime_type
ext = os.path.splitext(file_path)[1].lower()
mime_map = {
'.mp4': 'video/mp4',
'.avi': 'video/avi',
'.mov': 'video/quicktime',
'.mkv': 'video/x-matroska',
'.webm': 'video/webm',
'.jpg': 'image/jpeg',
'.jpeg': 'image/jpeg',
'.png': 'image/png',
'.gif': 'image/gif',
'.webp': 'image/webp'
}
return mime_map.get(ext, 'image/jpeg')
def build_message_with_media(text, media_path=None, use_vision=False, service_model_name="qwen3-vl-235b"):
"""Determine whether to include media files and message format based on use_vision parameter and service_model_name"""
content = [{"type": "text", "text": text}]
if use_vision and media_path:
media_type = "video"
mime_type = get_media_mime_type(media_path)
media_url = media_path
if (
not is_url(media_path)
and media_type == "video"
and service_model_name not in {"kimi-k2.6", "gemini-3.1-pro-preview"}
):
base64_media = encode_video(media_path)
if base64_media:
media_url = f"data:{mime_type};base64,{base64_media}"
# Build different message formats for different models
if service_model_name in ["glm-5v-turbo", "qwen3.6-plus", "mimo-v2.5-omni", "kimi-k2.6"]:
if media_type == 'video':
content.append({"type": "video_url", "video_url": {"url": media_url}})
else:
content.append({"type": "image_url", "image_url": {"url": media_path}})
elif service_model_name == "gemini-3.1-pro-preview":
# Gemini format
if is_url(media_path):
if media_type == 'video':
content.append({
"type": "input_audio",
"input_audio": {"data": media_path, "format": "video/mp4"}
})
else:
content.append({"type": "image_url", "image_url": {"url": media_path}})
else:
if media_type == 'video':
base64_media = encode_video(media_path)
if base64_media:
content.append({
"type": "input_audio",
"input_audio": {"data": base64_media, "format": mime_type}
})
else:
base64_media = encode_image(media_path)
if base64_media:
content.append({"type": "image_url", "image_url": {"url": f"{mime_type};base64,{base64_media}"}})
else:
# Default format
if media_type == 'video':
content.append({"type": "video_url", "video_url": {"url": media_url}})
else:
content.append({"type": "image_url", "image_url": {"url": media_path}})
return content
def build_message_with_image(text, image_path=None, use_vision=False, service_model_name="qwen3-vl-235b"):
"""Backward compatible function, recommend using build_message_with_media"""
return build_message_with_media(text, image_path, use_vision, service_model_name)
# --- Tool execution functions ---
def execute_tool(db_instance, tool_calls_data):
"""Parse and execute one or more tool calls, return structured results containing role, tool_name, parameters, content"""
if isinstance(tool_calls_data, dict):
tool_calls_data = [tool_calls_data]
elif not isinstance(tool_calls_data, list):
return [{"role": "tool", "tool_name": "unknown", "parameters": {}, "content": json.dumps({"error": "Invalid tool call format. Expected dict or list."}, ensure_ascii=False)}]
results = []
for tool_call_obj in tool_calls_data:
try:
method_name = tool_call_obj.get("tool_name")
if not method_name:
results.append({
"role": "tool",
"tool_name": "unknown",
"parameters": {},
"content": json.dumps({"error": "Missing tool identifier 'tool_name'"}, ensure_ascii=False)
})
continue
params = tool_call_obj.get("parameters", tool_call_obj.get("arguments", {}))
print(f" [Tool Execution] Calling: {method_name} Parameters: {params}")
if hasattr(db_instance, method_name):
method = getattr(db_instance, method_name)
result = method(**params)
print(f" [Tool Execution] Return result: {result}")
results.append({
"role": "tool",
"tool_name": method_name,
"parameters": params,
"content": json.dumps(result, ensure_ascii=False, default=str)
})
else:
results.append({
"role": "tool",
"tool_name": method_name,
"parameters": params,
"content": json.dumps({"error": f"Tool '{method_name}' not found"}, ensure_ascii=False)
})
except Exception as e:
results.append({
"role": "tool",
"tool_name": tool_call_obj.get("tool_name") or "unknown",
"parameters": tool_call_obj.get("parameters", tool_call_obj.get("arguments", {})),
"content": json.dumps({"error": str(e)}, ensure_ascii=False)
})
return results
# --- Tool call check functions ---
def check_tool_call(response_text):
"""
Extract all JSON format strings containing tool calls from text
Recognition rule: a JSON object must contain the top-level tool_name key.
Returns: (whether tool call was found, extracted tool call JSON list)
"""
text = response_text.strip()
# Only ``tool_name`` identifies a tool. ``name`` is also a common business
# parameter (for example, a household item name), so accepting it here can
# turn a nested parameters object into a false tool call.
def is_tool_call(obj):
return isinstance(obj, dict) and "tool_name" in obj
# Try parsing entire text as JSON first
try:
data = json.loads(text)
if isinstance(data, list) and len(data) > 0:
tool_calls = [item for item in data if is_tool_call(item)]
if tool_calls:
return True, tool_calls
elif is_tool_call(data):
return True, [data]
except json.JSONDecodeError:
pass
# Regex to extract all possible JSON object fragments
json_pattern = r'\{([^{}]|(\{[^{}]*\}))*\}'
potential_jsons = re.findall(json_pattern, text, re.DOTALL)
valid_tool_calls = []
for match in potential_jsons:
json_str = match[0] if isinstance(match, tuple) else match
if not json_str.startswith('{'):
json_str = '{' + json_str
if not json_str.endswith('}'):
json_str = json_str + '}'
try:
obj = json.loads(json_str)
if is_tool_call(obj):
valid_tool_calls.append(obj)
except (json.JSONDecodeError, ValueError):
continue
# Additional handling for JSON array scenario
if not valid_tool_calls:
array_pattern = r'\[.*\]'
array_matches = re.findall(array_pattern, text, re.DOTALL)
for arr_str in array_matches:
try:
arr = json.loads(arr_str)
if isinstance(arr, list):
for item in arr:
if is_tool_call(item):
valid_tool_calls.append(item)
except (json.JSONDecodeError, ValueError):
continue
if valid_tool_calls:
return True, valid_tool_calls
return False, None
# --- User response evaluation functions ---
def _evaluate_user_response_internal(user_response, user_instruction, last_agent_response, history_str="", user_model_name="Qwen3.5-397B-A17B"):
"""Internal helper function: build prompt and call LLM to evaluate user response"""
eval_user_content = f"""
[User Original Instruction]
{user_instruction}
[Interaction Process]
{history_str}
[Service Agent Response]
{last_agent_response}
[Simulated User Response]
{user_response}
"""
contradiction_check_messages = [
{"role": "system", "content": USER_CONTRADICTION_CHECK_PROMPT},
{"role": "user", "content": eval_user_content}
]
eval_response_text, _, _ = call_llm(contradiction_check_messages, agent_type="user", user_model_name=user_model_name)
evaluation_result = {}
try:
import re
code_block_match = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", eval_response_text, re.DOTALL)
if code_block_match:
evaluation_result = json.loads(code_block_match.group(1))
else:
try:
evaluation_result = json.loads(eval_response_text)
except:
start = eval_response_text.find('{')
end = eval_response_text.rfind('}')
if start != -1 and end != -1:
json_str = eval_response_text[start:end+1]
evaluation_result = json.loads(json_str)
else:
return {"error": "No JSON found", "raw": eval_response_text}
except Exception as e:
return {"error": str(e), "raw": eval_response_text}
return evaluation_result
def check_user_contradiction(user_response, user_instruction, image_description="", multi_agent_user=False, last_agent_response="", history=[], summarized_history=None, user_mode="easy", user_model_name="Qwen3.5-397B-A17B", max_retries=3):
"""
Check if user response contradicts the task (based on scoring mechanism)
If multi_agent_user is True and there is a contradiction (any dimension <= 0 points), use LLM to correct user response
Will retry up to max_retries times until the evaluation passes
:param user_mode: User difficulty mode ("easy" or "hard"), used to select corresponding correction prompt
:param max_retries: Maximum number of retry attempts (default: 3)
:return: (corrected_response, evaluation_result)
"""
if not multi_agent_user:
return user_response, None
# Build history conversation string
if summarized_history:
history_str = summarized_history
else:
history_str = ""
for item in history:
role = "User" if item["role"] == "user" else "Service Agent"
content = item["content"]
history_str += f"{role}: {content}\n"
# Select corresponding prompt based on user_mode
if user_mode == "hard":
correction_prompt_template = USER_TEXT_ONLY_PROMPT_HARD
else:
correction_prompt_template = USER_TEXT_ONLY_PROMPT_EASY
current_response = user_response
all_attempts = []
# Evaluate original response first
print("Evaluating user response (Corrections Check)...")
evaluation_result = _evaluate_user_response_internal(current_response, user_instruction, last_agent_response, history_str, user_model_name="Qwen3.5-397B-A17B")
if "error" in evaluation_result:
print(f"Evaluation failed: {evaluation_result['error']}")
return user_response, evaluation_result
scores = evaluation_result.get("scores", {})
try:
score_values = [float(v) for v in scores.values()]
avg_score = round(sum(score_values) / len(score_values), 2) if score_values else 0
except:
avg_score = 0
evaluation_result["average_score"] = avg_score
# Check if correction is needed
def needs_correction_check(scores_dict):
for k, v in scores_dict.items():
try:
if float(v) <= 0:
return True
except:
pass
return False
needs_fix = needs_correction_check(scores)
corrected_response = current_response
evaluation_result["original_response"] = user_response
if not needs_fix:
evaluation_result["correction_applied"] = False
evaluation_result["retry_count"] = 0
print(f"Scores passed (average: {avg_score})")
return corrected_response, evaluation_result
# Need correction - start retry loop
print(f"Low score detected (average: {avg_score}), starting to correct user response... (Scores: {scores})")
for attempt in range(1, max_retries + 1):
print(f"\n--- Correction Attempt {attempt}/{max_retries} ---")
# Get feedback from previous attempt
if attempt == 1:
feedback = evaluation_result.get("suggestion", "")
response_to_correct = user_response
else:
feedback = all_attempts[-1].get("suggestion", "")
response_to_correct = all_attempts[-1].get("response", user_response)
correction_input = correction_prompt_template.format(
user_instruction=user_instruction,
image_description=image_description,
original_user_response=response_to_correct,
evaluation_feedback=feedback,
history_summary=history_str,
service_agent_response=last_agent_response
)
correction_messages = [{"role": "user", "content": correction_input}]
corrected_response_text, _, _ = call_llm(correction_messages, agent_type="user", user_model_name=user_model_name)
cleaned_response = corrected_response_text.strip()
if cleaned_response.startswith("```"):
lines = cleaned_response.split('\n')
if len(lines) >= 2:
cleaned_response = '\n'.join(lines[1:-1])
corrected_response = cleaned_response.strip()
print(f"Corrected response (attempt {attempt}): {corrected_response}")
# Evaluate the corrected response
print(f"Evaluating corrected response (attempt {attempt})...")
corrected_evaluation = _evaluate_user_response_internal(corrected_response, user_instruction, last_agent_response, history_str, user_model_name=user_model_name)
attempt_record = {
"attempt": attempt,
"response": corrected_response,
"scores": corrected_evaluation.get("scores", {}),
"suggestion": corrected_evaluation.get("suggestion", "")
}
if "error" not in corrected_evaluation:
corrected_scores = corrected_evaluation.get("scores", {})
try:
c_values = [float(v) for v in corrected_scores.values()]
c_avg_score = round(sum(c_values) / len(c_values), 2) if c_values else 0
except:
c_avg_score = 0
attempt_record["average_score"] = c_avg_score
print(f"Attempt {attempt} scores: {corrected_scores} (average: {c_avg_score})")
# Check if this attempt passes
if not needs_correction_check(corrected_scores):
print(f"Attempt {attempt} passed evaluation!")
all_attempts.append(attempt_record)
break
else:
print(f"Attempt {attempt} still has low scores, will retry...")
else:
attempt_record["error"] = corrected_evaluation.get("error")
print(f"Attempt {attempt} evaluation failed: {corrected_evaluation.get('error')}")
all_attempts.append(attempt_record)
# If this is the last attempt, we keep the corrected_response from this attempt
if attempt == max_retries:
print(f"Reached maximum retries ({max_retries}), using last attempt's response")
# Build final evaluation result
evaluation_result["correction_applied"] = True
evaluation_result["retry_count"] = len(all_attempts)
evaluation_result["corrected_response"] = corrected_response
evaluation_result["all_attempts"] = all_attempts
if all_attempts:
final_attempt = all_attempts[-1]
evaluation_result["corrected_scores"] = final_attempt.get("scores", {})
evaluation_result["corrected_average_score"] = final_attempt.get("average_score", 0)
return corrected_response, evaluation_result