SQLAssistant-MCP Evaluation Scripts
Evaluation tools for analyzing SQLAssistant-MCP multi-agent system execution results across different LLM models.
Scripts
analyze_retry_patterns.py
Analyzes RETRY patterns in execution traces, including:
- Orchestrator-level retries (exception handling)
- Business logic retries (compliance check failures)
Output: retry_summary.csv, retry_analysis.json, error_by_agent.csv, business_retry_chapters.csv
evaluate_scores.py
Collects score statistics from execution_log.json files.
Output: score_summary.csv, score_analysis.json
evaluate_success.py
Measures success rate based on whether get_database_schema tool was called (indicates proper database exploration vs hallucination).
Output: success_rate.csv, success_detailed_results.json
evaluate_trajectory-Filter_Tools.py
Evaluates trajectory metrics against reference trajectory:
- Exact match
- In-order match
- Any-order match
- Precision / Recall
- Single-tool use
- Path diversity metrics
Output: evaluation_results.csv, any_order_match_failures.csv
Configuration
reference_trajectory.yaml defines:
- Reference trajectory (ground truth)
- Target tools for evaluation
- Models to evaluate
- Extraction types (SPAN, Chain, Agent, LLM, Tool)
Usage
python analyze_retry_patterns.py
python evaluate_scores.py
python evaluate_success.py
python evaluate_trajectory-Filter_Tools.py --config reference_trajectory.yaml
Requirements
- Python 3.8+
- pandas
- pyyaml