data/processed/RQ1/RecruitmentAssistant-H_A2A/README.md

RecruitmentAssistant-H_A2A Evaluation Scripts

This directory contains evaluation scripts for analyzing the RecruitmentAssistant-H_A2A (Hybrid A2A) project results.

Architecture

This project uses a hybrid Agent-to-Agent (A2A) architecture:

• LangGraph - Job Analysis stage with dynamic batched tool execution
• CrewAI - Candidate Evaluation stage with standard CrewAI structure
• AutoGen - Interview Communication stage with flexible LLM/tool patterns

Scripts

1. analyze_retry_patterns.py

Analyzes RETRY patterns in execution paths.

Features:

• Counts error occurrences by location and type
• Calculates RETRY rates per model
• Identifies maximum RETRY attempts
• Tracks errors by agent and node type

Output:

• retry_analysis.json - Detailed JSON results
• retry_summary.csv - Summary statistics
• error_by_agent.csv - Error breakdown by agent

Usage:

python3 analyze_retry_patterns.py

2. evaluate_success.py

Calculates success rates for each model.

Success Criteria:

• Reports folder exists
• Contains at least 2 markdown files

Output:

• success_detailed_results.json - Detailed results per session
• success_rate.csv - Success rate summary

Usage:

python3 evaluate_success.py

3. evaluate_trajectory-mix.py

Evaluates trajectory metrics with hybrid architecture support.

Special Features:

• Dynamic LangGraph matching - Handles batched tool execution patterns (3-4 searches)
• AutoGen LLM insertion - Flexible LLM call patterns between tools
• Tool order permutation - Evaluates different tool call orders

Metrics:

• Exact match
• In-order match
• Any-order match
• Precision
• Recall
• Single-tool use
• Path diversity (unique path ratio, path entropy)

Output:

• evaluation_results.csv - Trajectory evaluation results

Usage:

python3 evaluate_trajectory-mix.py --config reference_trajectory.yaml

4. reference_trajectory.yaml

Configuration file defining reference trajectories and evaluation parameters.

Key Sections:

• reference_trajectory_3x - Reference for 3 web searches
• reference_trajectory_4x - Reference for 4 web searches
• target_tools - Required tools for evaluation
• models - List of models to evaluate
• permutable_tool_groups - Tool groups that can appear in any order
• extract_types - Node types to extract (SPAN, Chain, AGENT, LLM, Tool)

Dynamic Matching

The hybrid architecture requires sophisticated matching:

Layer 1: LangGraph Tools Batching

• 3 searches: 4 grouping patterns [1,1,1], [1,2], [2,1], [3]
• 4 searches: 7 grouping patterns [1,1,1,1], [2,1,1], [1,2,1], [1,1,2], [3,1], [1,3], [4]

Layer 2: AutoGen LLM Insertion

• 2 tools: 2 patterns (0 or 1 LLM between tools)
• 3 tools: 4 patterns (2^2 combinations)

Layer 3: Tool Order Permutation

• Candidate evaluation tools can appear in any order
• Interview communication tools can appear in any order

Total combinations: Up to 7 × 2 × (permutations) variants per sample

Configuration

Edit reference_trajectory.yaml to:

• Modify reference trajectories
• Add/remove models
• Adjust tool evaluation criteria
• Configure dynamic matching patterns
• Enable/disable pattern matching layers

Requirements

• Python 3.7+
• pandas
• pyyaml

Install dependencies:

pip install pandas pyyaml

Data Structure

Expected directory structure:

/Users/wzr/TOSEM-2025/RESULTS/
├── ModelName/
│   └── RecruitmentAssistant-H_A2A/
│       └── test_results/
│           ├── session_1/
│           │   ├── execution_path.md
│           │   └── reports/
│           │       ├── file1.md
│           │       └── file2.md
│           └── session_2/
│               └── ...

Notes

• The hybrid architecture requires more flexible matching than standard frameworks
• Dynamic reference selection adapts to different model execution strategies
• All scripts output CSV files for easy analysis
• JSON files contain detailed per-session information
• Modify BASE_DIR in scripts if using a different results directory
• The evaluation automatically selects the best-matching reference variant per sample