王子睿
restructure + add files
8c10cf2
Raw
History Blame Contribute Delete
29.9 kB
#!/usr/bin/env python3
"""
Trajectory Evaluation Script - EmailResponder Project
Evaluates 6 trajectory metrics:
1. Exact match
2. In-order match
3. Any-order match
4. Precision
5. Recall
6. Single-tool use
"""
import os
import re
import yaml
from pathlib import Path
from typing import List, Dict, Tuple, Set
from collections import defaultdict
import pandas as pd
import math
from itertools import permutations, product
class TrajectoryParser:
"""Parse execution_path.md file and extract complete execution trajectory"""
def __init__(self, md_file_path: str, extract_types: List[str] = None):
"""
Args:
md_file_path: execution_path.md file path
extract_types: List of node types to extract, default ['Tool']
Available types: 'SPAN', 'Chain', 'Tool', 'AGENT', 'LLM', 'Task Created', 'Crew Created'
"""
self.md_file_path = md_file_path
self.extract_types = extract_types or ["Tool"]
self.trajectory = []
def parse(self) -> List[str]:
"""Parse file and return execution trajectory sequence"""
if not os.path.exists(self.md_file_path):
return []
with open(self.md_file_path, "r", encoding="utf-8") as f:
content = f.read()
# Extract Execution Path Tree section (in code block)
tree_match = re.search(
r"## Execution Path Tree.*?```\n(.*?)```", content, re.DOTALL
)
if not tree_match:
return []
tree_content = tree_match.group(1)
trajectory = []
for line in tree_content.split("\n"):
# Remove tree structure characters, keep node content
clean_line = re.sub(r"^[│├└─\s]+", "", line).strip()
if not clean_line:
continue
# Remove error marker ❌
clean_line = re.sub(r"^❌\s+", "", clean_line)
# Remove RETRY markers (retry N) and [RETRYN]
clean_line = re.sub(r"\s*\(retry\s+\d+\)", "", clean_line)
clean_line = re.sub(r"\s*\[RETRY\d+\]", "", clean_line)
# Remove ERROR info (keep node type and name, remove [ERROR:...] part)
clean_line = re.sub(r"\s*\[ERROR:[^\]]*\]", "", clean_line)
# Extract node info
node_info = self._extract_node_info(clean_line)
if node_info and node_info["type"] in self.extract_types:
trajectory.append(node_info["action"])
self.trajectory = trajectory
return trajectory
def _extract_node_info(self, line: str) -> dict:
"""Extract node information from a line
Note: Input line should already have error markers ❌, RETRY markers, and ERROR info removed
Returns:
{'type': str, 'action': str} or None
"""
# SPAN node: [SPAN] span_name [statistics]
# More lenient matching, handling possible residual special characters
span_match = re.match(r"\[SPAN\]\s+([^\[]+?)(?:\s+\[.*?\])*\s*$", line)
if span_match:
span_name = span_match.group(1).strip()
return {"type": "SPAN", "action": f"SPAN: {span_name}"}
# Chain node: [Chain] chain_name [statistics]
# For Crew_xxx.kickoff format, use wildcard Crew***.kickoff
chain_match = re.match(r"\[Chain\]\s+([^\[]+?)(?:\s+\[.*?\])*\s*$", line)
if chain_match:
chain_name = chain_match.group(1).strip()
# Wildcard handling: Crew_UUID.kickoff -> Crew***.kickoff
chain_name = re.sub(
r"Crew_[a-f0-9\-]+\.kickoff", "Crew***.kickoff", chain_name
)
return {"type": "Chain", "action": f"Chain: {chain_name}"}
# AGENT node: [AGENT] agent_name._execute_core [statistics]
agent_match = re.match(r"\[AGENT\]\s+([^\[]+?)(?:\s+\[.*?\])*\s*$", line)
if agent_match:
agent_name = agent_match.group(1).strip()
# Remove ._execute_core suffix
agent_name = re.sub(r"\._execute_core$", "", agent_name)
return {"type": "AGENT", "action": f"Agent: {agent_name}"}
# Tool node: [Tool] tool_name._use [time]
tool_match = re.match(
r"\[Tool\]\s+([^\[\]]+?)(?:\s+\[[\d.]+(?:ms|s)\])?(?:\s*@@@)?\s*$", line
)
if tool_match:
tool_name = tool_match.group(1).strip()
# Remove ._use suffix
tool_name = re.sub(r"\._use$", "", tool_name)
return {"type": "Tool", "action": f"Tool: {tool_name}"}
# LLM node: [LLM] model_name (tokens) [time]
llm_match = re.match(r"\[LLM\]\s+([^\(\[]+)", line)
if llm_match:
model_name = llm_match.group(1).strip()
return {"type": "LLM", "action": f"LLM: {model_name}"}
# Task Created node: [Task Created] [time]
task_match = re.match(r"\[Task Created\]", line)
if task_match:
return {"type": "Task Created", "action": "Task Created"}
# Crew Created node: [Crew Created] [time]
crew_match = re.match(r"\[Crew Created\]", line)
if crew_match:
return {"type": "Crew Created", "action": "Crew Created"}
return None
class TrajectoryEvaluator:
"""Trajectory Evaluator - Implements 6 evaluation metrics"""
def __init__(self, reference_trajectory: List[str]):
"""
Args:
reference_trajectory: Reference trajectory (ground truth)
"""
self.reference = reference_trajectory
def _match_action(self, predicted_action: str, reference_action: str) -> bool:
"""
Match two actions, supports wildcards
Args:
predicted_action: Actually executed action
reference_action: Reference action (may contain wildcards)
Returns:
True if match, False otherwise
"""
# Exact match
if predicted_action == reference_action:
return True
# Wildcard match: LLM: * matches any LLM: <model_name>
if reference_action == "LLM: *" and predicted_action.startswith("LLM: "):
return True
return False
def exact_match(self, predicted: List[str]) -> int:
"""
Exact match: Predicted trajectory must be exactly the same as reference (supports wildcards)
Returns:
1 if exact match, 0 otherwise
"""
if len(predicted) != len(self.reference):
return 0
for i in range(len(predicted)):
if not self._match_action(predicted[i], self.reference[i]):
return 0
return 1
def in_order_match(self, predicted: List[str]) -> int:
"""
In-order match: Reference trajectory must be a subsequence of predicted (supports wildcards)
Allows extra actions, but core steps must appear in order
Returns:
1 if in-order match, 0 otherwise
"""
if not self.reference:
return 1 # Empty reference always matches
ref_idx = 0
for pred_action in predicted:
if ref_idx < len(self.reference) and self._match_action(
pred_action, self.reference[ref_idx]
):
ref_idx += 1
# Check if all reference steps were found in order
return 1 if ref_idx == len(self.reference) else 0
def any_order_match(self, predicted: List[str]) -> int:
"""
Any-order match: As long as predicted contains all necessary actions (supports wildcards)
Order doesn't matter, allows extra actions
Returns:
1 if any-order match, 0 otherwise
"""
if not self.reference:
return 1
# Create copy of predicted actions for matching
pred_remaining = predicted.copy()
# For each reference action, try to find a match in predicted
for ref_action in self.reference:
matched = False
for i, pred_action in enumerate(pred_remaining):
if self._match_action(pred_action, ref_action):
pred_remaining.pop(i) # Remove matched
matched = True
break
if not matched:
return 0 # Reference action not found
return 1
def precision(self, predicted: List[str]) -> float:
"""
Precision: How many in predicted trajectory are considered correct by reference (supports wildcards)
Precision = TP / (TP + FP)
TP: Correct tool calls in prediction
FP: Incorrect/extra tool calls in prediction
Returns:
precision value (0.0 - 1.0)
"""
if not predicted:
return 1.0 # No prediction, no false prediction
if not self.reference:
return 0.0 # Reference is empty but has prediction, all wrong
# Create copy of reference actions for matching
ref_remaining = self.reference.copy()
tp = 0 # True Positives
for pred_action in predicted:
# Try to find match in reference
for i, ref_action in enumerate(ref_remaining):
if self._match_action(pred_action, ref_action):
tp += 1
ref_remaining.pop(i) # Remove matched
break
fp = len(predicted) - tp # False Positives
return tp / (tp + fp) if (tp + fp) > 0 else 0.0
def recall(self, predicted: List[str]) -> float:
"""
Recall: How many in reference trajectory are covered by predicted (supports wildcards)
Recall = TP / (TP + FN)
TP: Necessary calls covered by prediction
FN: Missed necessary calls
Returns:
recall value (0.0 - 1.0)
"""
if not self.reference:
return 1.0 # Reference is empty, nothing to recall
if not predicted:
return 0.0 # No prediction, recall is 0
# Create copy of predicted actions for matching
pred_remaining = predicted.copy()
tp = 0 # True Positives
for ref_action in self.reference:
# Try to find match in predicted
for i, pred_action in enumerate(pred_remaining):
if self._match_action(pred_action, ref_action):
tp += 1
pred_remaining.pop(i) # Remove matched
break
fn = len(self.reference) - tp # False Negatives
return tp / (tp + fn) if (tp + fn) > 0 else 0.0
def single_tool_use(self, predicted: List[str], tool_name: str) -> int:
"""
Single-tool use: Check if specific tool appears in trajectory (supports wildcards)
Args:
predicted: Predicted trajectory
tool_name: Target tool name
Returns:
1 if tool is used, 0 otherwise
"""
# Check if any action in predicted trajectory matches target tool
for pred_action in predicted:
if self._match_action(pred_action, tool_name):
return 1
return 0
def evaluate_all(
self, predicted: List[str], target_tools: List[str] = None
) -> Dict[str, float]:
"""
Evaluate all metrics
Args:
predicted: Predicted trajectory
target_tools: List of tools to check usage (for single-tool use)
Returns:
Dictionary of all metric evaluation results
"""
results = {
"exact_match": self.exact_match(predicted),
"in_order_match": self.in_order_match(predicted),
"any_order_match": self.any_order_match(predicted),
"precision": self.precision(predicted),
"recall": self.recall(predicted),
}
# Single-tool use metric - calculate overall usage rate (average of all tools)
if target_tools:
tool_usage_count = sum(
self.single_tool_use(predicted, tool) for tool in target_tools
)
results["single_tool_use"] = (
tool_usage_count / len(target_tools) if target_tools else 0.0
)
return results
class DatasetEvaluator:
"""Dataset-level evaluator"""
def __init__(self, config_file: str):
"""
Args:
config_file: YAML config file path containing reference trajectory definition
"""
self.config_file = config_file
self.config = self._load_config()
self.reference_trajectory = self.config.get("reference_trajectory", [])
# Dynamic reference trajectory: select based on unified_web_search count
self.reference_trajectory_3x = self.config.get("reference_trajectory_3x", None)
self.reference_trajectory_4x = self.config.get("reference_trajectory_4x", None)
self.use_dynamic_reference = (
self.reference_trajectory_3x is not None
and self.reference_trajectory_4x is not None
)
self.target_tools = self.config.get("target_tools", [])
self.models = self.config.get("models", [])
self.project_name = self.config.get("project_name", "EmailResponder")
# Trajectory extraction types: default only extract Tool, can also configure as ['Tool', 'AGENT', 'Task Created'] etc.
self.extract_types = self.config.get("extract_types", ["Tool"])
# Permutable tool groups configuration (for dynamic tool permutation optimization)
self.permutable_tool_groups = self.config.get("permutable_tool_groups", {})
def _load_config(self) -> Dict:
"""Load YAML config file"""
if not os.path.exists(self.config_file):
print(f"⚠️ Config file not found: {self.config_file}")
return {}
with open(self.config_file, "r", encoding="utf-8") as f:
return yaml.safe_load(f)
def _count_unified_web_search_in_analyze_job(self, exec_path_file: str) -> int:
"""
Count unified_web_search occurrences under analyze_job SPAN
Args:
exec_path_file: execution_path.md file path
Returns:
unified_web_search call count
"""
if not os.path.exists(exec_path_file):
return 0
with open(exec_path_file, "r", encoding="utf-8") as f:
content = f.read()
# Extract Execution Path Tree
tree_match = re.search(
r"## Execution Path Tree.*?```\n(.*?)```", content, re.DOTALL
)
if not tree_match:
return 0
tree_content = tree_match.group(1)
lines = tree_content.split("\n")
# Find analyze_job SPAN range
in_analyze_job = False
analyze_job_level = -1
count = 0
for line in lines:
# Calculate current line level (by leading tree characters)
level = len(re.match(r"^([│├└─\s]*)", line).group(1))
# Clean line content
clean_line = re.sub(r"^[│├└─\s]+", "", line).strip()
clean_line = re.sub(r"^❌\s+", "", clean_line)
# Check if entering analyze_job SPAN
if "[SPAN] analyze_job" in clean_line:
in_analyze_job = True
analyze_job_level = level
continue
# If in analyze_job SPAN
if in_analyze_job:
# If encountering same or higher level SPAN, left analyze_job
if "[SPAN]" in clean_line and level <= analyze_job_level:
break
# Count unified_web_search
if "[Tool] unified_web_search" in clean_line:
count += 1
return count
def _generate_permuted_trajectories(
self, base_trajectory: List[str]
) -> List[List[str]]:
"""
Generate all possible tool permutation trajectories based on permutable_tool_groups config
Args:
base_trajectory: Base reference trajectory
Returns:
List of all possible permutation trajectories (including original)
"""
if not self.permutable_tool_groups:
# No permutable tool groups configured, return original trajectory
return [base_trajectory]
# Collect all permutable tool groups and their positions in trajectory
tool_groups_positions = []
for group_name, tools in self.permutable_tool_groups.items():
# Find positions of this tool group in trajectory
positions = []
tool_indices = {}
for i, action in enumerate(base_trajectory):
for tool in tools:
if action == tool:
positions.append(i)
tool_indices[i] = tool
break
# Only permute when all tools are found
if len(positions) == len(tools):
# Record positions and tools for this group
tools_at_positions = [tool_indices[pos] for pos in positions]
tool_groups_positions.append((positions, tools_at_positions))
if not tool_groups_positions:
# No complete permutable tool groups found
return [base_trajectory]
# Generate all possible permutation combinations
all_trajectories = []
# Generate all permutations for each tool group
group_permutations = []
for positions, tools in tool_groups_positions:
# Generate all permutations for this tool group
perms = list(permutations(tools))
group_permutations.append([(positions, perm) for perm in perms])
# Cartesian product: combine all tool group permutations
all_group_combinations = list(product(*group_permutations))
# Generate a new trajectory for each combination
for combination in all_group_combinations:
new_trajectory = base_trajectory.copy()
# Apply all tool permutations in this combination
for positions, perm in combination:
for pos, tool in zip(positions, perm):
new_trajectory[pos] = tool
all_trajectories.append(new_trajectory)
return all_trajectories
def _find_best_reference_trajectory(
self, predicted: List[str], candidate_references: List[List[str]]
) -> Tuple[List[str], Dict[str, float]]:
"""
Select the best reference trajectory from multiple candidates
Args:
predicted: Predicted trajectory
candidate_references: Candidate reference trajectory list
Returns:
(Best reference trajectory, corresponding match score dict)
"""
best_reference = candidate_references[0]
best_score = -1
best_metrics = {}
for ref_trajectory in candidate_references:
evaluator = TrajectoryEvaluator(ref_trajectory)
# Calculate three key matching metrics
exact = evaluator.exact_match(predicted)
in_order = evaluator.in_order_match(predicted)
any_order = evaluator.any_order_match(predicted)
# Composite score: exact_match has highest weight, followed by in_order_match
# Use weighted sum: exact*3 + in_order*2 + any_order*1
score = exact * 3 + in_order * 2 + any_order * 1
if score > best_score:
best_score = score
best_reference = ref_trajectory
best_metrics = {
"exact_match": exact,
"in_order_match": in_order,
"any_order_match": any_order,
}
return best_reference, best_metrics
def collect_execution_paths(
self, model_name: str, base_dir: str
) -> List[Tuple[str, List[str]]]:
"""
Collect and parse all execution_path.md files for specified model
Args:
model_name: Model name
base_dir: RESULTS directory path
Returns:
[(session_id, trajectory), ...] list
"""
model_dir = Path(base_dir) / model_name / self.project_name / "test_results"
if not model_dir.exists():
print(f"⚠️ Model directory not found: {model_dir}")
return []
results = []
# Iterate through all session subdirectories
for session_dir in sorted(model_dir.iterdir()):
if not session_dir.is_dir():
continue
exec_path_file = session_dir / "execution_path.md"
if not exec_path_file.exists():
continue
# Parse trajectory (using configured extraction types)
parser = TrajectoryParser(
str(exec_path_file), extract_types=self.extract_types
)
trajectory = parser.parse()
results.append((session_dir.name, trajectory))
return results
def evaluate_model(self, model_name: str, base_dir: str) -> Dict[str, float]:
"""
Evaluate single model's performance on all samples
Args:
model_name: Model name
base_dir: RESULTS directory path
Returns:
Average metrics dictionary
"""
model_dir = Path(base_dir) / model_name / self.project_name / "test_results"
if not model_dir.exists():
print(f"⚠️ Model {model_name} has no trajectory data found")
return {}
# Accumulate metrics for all samples
all_metrics = defaultdict(list)
trajectories = []
# Iterate through all sessions
for session_dir in sorted(model_dir.iterdir()):
if not session_dir.is_dir():
continue
exec_path_file = session_dir / "execution_path.md"
if not exec_path_file.exists():
continue
# Parse trajectory
parser = TrajectoryParser(
str(exec_path_file), extract_types=self.extract_types
)
predicted = parser.parse()
# Dynamically select reference trajectory
if self.use_dynamic_reference:
# Count unified_web_search occurrences in analyze_job
search_count = self._count_unified_web_search_in_analyze_job(
str(exec_path_file)
)
# Select reference trajectory based on count
if search_count <= 3:
reference = self.reference_trajectory_3x
else:
reference = self.reference_trajectory_4x
else:
# Use default reference trajectory
reference = self.reference_trajectory
# Dynamic tool permutation optimization
if self.permutable_tool_groups:
# Generate all possible tool permutation trajectories
candidate_references = self._generate_permuted_trajectories(reference)
# Select the best from all candidate reference trajectories
if len(candidate_references) > 1:
reference, _ = self._find_best_reference_trajectory(
predicted, candidate_references
)
# else: keep reference unchanged
# Create evaluator and evaluate
evaluator = TrajectoryEvaluator(reference)
metrics = evaluator.evaluate_all(predicted, self.target_tools)
for key, value in metrics.items():
all_metrics[key].append(value)
trajectories.append((session_dir.name, predicted))
# Calculate averages
avg_metrics = {}
for key, values in all_metrics.items():
avg_metrics[key] = sum(values) / len(values) if values else 0.0
num_samples = len(trajectories)
if num_samples > 0:
path_counter = defaultdict(int)
for session_id, predicted in trajectories:
path_key = tuple(predicted)
path_counter[path_key] += 1
unique_paths = len(path_counter)
unique_path_ratio = unique_paths / num_samples if num_samples > 0 else 0.0
probs = [count / num_samples for count in path_counter.values()]
H = -sum(p * math.log(p) for p in probs if p > 0)
if len(probs) > 1:
path_entropy = H / math.log(len(probs))
else:
path_entropy = 0.0
else:
unique_path_ratio = 0.0
path_entropy = 0.0
avg_metrics["unique_path_ratio"] = unique_path_ratio
avg_metrics["path_entropy"] = path_entropy
# Add sample count
avg_metrics["num_samples"] = num_samples
return avg_metrics
def evaluate_all_models(self, base_dir: str = None) -> pd.DataFrame:
"""
Evaluate all models and generate summary table
Args:
base_dir: RESULTS directory path, defaults to two levels up from script directory
Returns:
DataFrame containing all model evaluation results
"""
if base_dir is None:
# Default path: two levels up from script directory
base_dir = Path(__file__).parent.parent.parent
results = []
for model_name in self.models:
print(f"\n📊 Evaluating model: {model_name}")
metrics = self.evaluate_model(model_name, str(base_dir))
if metrics:
metrics["model"] = model_name
results.append(metrics)
print(f" ✅ Completed, sample count: {metrics['num_samples']}")
else:
print(f" ❌ Skipped (no data)")
if not results:
print("\n❌ No model data available")
return pd.DataFrame()
# Create DataFrame
df = pd.DataFrame(results)
# Adjust column order: model name first
cols = [
"model",
"num_samples",
"exact_match",
"in_order_match",
"any_order_match",
"precision",
"recall",
"single_tool_use",
"unique_path_ratio",
"path_entropy",
]
# Keep only existing columns
cols = [col for col in cols if col in df.columns]
df = df[cols]
return df
def main():
"""Main function"""
import argparse
parser = argparse.ArgumentParser(
description="Evaluate trajectory metrics for EmailResponder project"
)
parser.add_argument(
"--config",
type=str,
default="reference_trajectory.yaml",
help="Reference trajectory config file path (YAML format)",
)
parser.add_argument(
"--base-dir",
type=str,
default=None,
help="RESULTS directory path (defaults to two levels up from script directory)",
)
parser.add_argument(
"--output",
type=str,
default="evaluation_results.csv",
help="Output CSV file path",
)
args = parser.parse_args()
# If config not specified, use config file in script directory
config_path = args.config
if not os.path.isabs(config_path):
config_path = os.path.join(os.path.dirname(__file__), config_path)
print("=" * 80)
print("Trajectory Evaluation Tool - EmailResponder")
print("=" * 80)
print(f"\n📁 Config file: {config_path}")
# Create evaluator
evaluator = DatasetEvaluator(config_path)
print(f"📋 Project name: {evaluator.project_name}")
# Display reference trajectory info
if evaluator.use_dynamic_reference:
print(f"🎯 Reference trajectory: Dynamic selection")
print(
f" - 3x version (unified_web_search<=3): {len(evaluator.reference_trajectory_3x)} steps"
)
print(
f" - 4x version (unified_web_search>=4): {len(evaluator.reference_trajectory_4x)} steps"
)
else:
print(f"🎯 Reference trajectory: {evaluator.reference_trajectory}")
print(f"🔧 Target tools: {evaluator.target_tools}")
print(f"🤖 Evaluation models: {evaluator.models}")
# Display tool permutation info
if evaluator.permutable_tool_groups:
print(f"\n🔀 Dynamic tool permutation optimization: Enabled")
total_permutations = 1
for group_name, tools in evaluator.permutable_tool_groups.items():
num_perms = math.factorial(len(tools))
total_permutations *= num_perms
print(f" - {group_name}: {len(tools)} tools, {num_perms} permutations")
print(f" - Total permutation combinations: {total_permutations}")
print(
f" - Evaluation strategy: After dynamic 3x/4x selection, choose permutation with highest tool order match"
)
else:
print(f"\n🔀 Dynamic tool permutation optimization: Disabled")
# Evaluate all models
df = evaluator.evaluate_all_models(args.base_dir)
if df.empty:
print("\n❌ Evaluation failed: No data")
return
print("\n" + "=" * 80)
print("📊 Evaluation Results Summary")
print("=" * 80)
# Format display
pd.set_option("display.max_columns", None)
pd.set_option("display.width", None)
pd.set_option("display.float_format", lambda x: f"{x:.4f}")
print("\n" + df.to_string(index=False))
# Save results
output_dir = os.path.dirname(args.output) or "."
os.makedirs(output_dir, exist_ok=True)
csv_file = args.output
df.to_csv(csv_file, index=False)
print(f"\n✅ CSV file saved: {csv_file}")
print("\n" + "=" * 80)
print("✅ Evaluation completed")
print("=" * 80)
if __name__ == "__main__":
main()