AINativeBench / data /processed /RQ1 /MarkdownValidator-MCP /reference_trajectory.yaml
王子睿
restructure + add files
8c10cf2
Raw
History Blame Contribute Delete
7.84 kB
# Complete Reference Trajectory Configuration - MarkdownValidator-MCP Project
#
# Version: Ideal trajectory based on code design (applicable to all models)
# Principle: Only includes steps explicitly required by code design, not based on test statistics
#
# Code Design Analysis:
# 1. Single Agent executing single Task:
# - Requirements Manager: Uses markdown_validation_tool (MCP tool) to analyze Markdown file and generate JSON report
#
# 2. Tool Usage Analysis:
# - markdown_validation_tool provided through MCP server
# - Tool connects to MCP server via MCPServerAdapter
# - tasks.yaml requires calling this tool to validate file
# - Tool invocation is necessary step to complete task
#
# 3. Execution Flow:
# - Process.sequential: Sequential execution (though only one Task)
# - Agent executes Task, uses LLM to decide tool invocation, then uses LLM to generate final JSON
#
# 4. MCP Version Characteristics:
# - Tools obtained from MCP server via MCPServerAdapter
# - main.py has two nested langfuse spans (code duplication), resulting in two SPAN: markdown_validation layers
# - This is the actual execution structure, needs to be reflected in reference trajectory
# Project name
project_name: "MarkdownValidator-MCP"
# Trajectory extraction type configuration
# Per user requirements, focus on: SPAN, Chain, AGENT, LLM, Tool
# Ignore: Task Created, Crew Created and other non-core content
extract_types:
- "SPAN" # Level 1 and 2: Two nested markdown_validation SPANs
- "Chain" # Level 3: Crew execution chain
- "AGENT" # Level 4: Agent execution
- "LLM" # Level 5: LLM invocation
- "Tool" # Level 5: Tool invocation
# ============================================================================
# Reference Trajectory (Ground Truth) - Based on Code Design
# ============================================================================
#
# Graphical Structure (Ideal Execution Path):
#
# [SPAN] markdown_validation
# └─ [Chain] Crew***.kickoff
# └─ [AGENT] Requirements Manager ← Task: syntax_review_task
# ├─ [LLM] * ← Decide to use tool for validation
# ├─ [Tool] markdown_validation_tool ← Call MCP tool to validate file
# └─ [LLM] * ← Process validation results and generate JSON output
#
# Design Rationale:
# - crew.py: Defines 1 Agent (Requirements_Manager) and 1 Task (syntax_review_task)
# - crew.py: Agent uses self.mcp_adapter.tools to get MCP tools
# - tasks.yaml: Explicitly requires using markdown_validation_tool and outputting JSON format report
# - main.py line 280-281: Two nested `start_as_current_span(name="markdown_validation")`
# - agents.yaml: Agent role is "Requirements Manager"
# - MCP tools obtained from server via MCPServerAdapter
#
# Ideal Trajectory Explanation:
# - Single markdown_validation SPAN (nested second layer filtered by evaluation script)
# - Requirements Manager: 2 LLM calls + 1 tool call
# 1. First LLM: Analyze task, decide to use MCP tool
# 2. Tool call: markdown_validation_tool validates file (via MCP)
# 3. Second LLM: Process tool results, generate JSON format report
# - This trajectory represents ideal, no-retry, no-redundancy execution path
# - Note: Actual execution may have nested SPANs, evaluation script auto-filters for consistency
#
# Notes:
# - "LLM: *" means any LLM model (wildcard match)
# - Agent name is "Requirements Manager" (role field, with space)
# - Tool name is "markdown_validation_tool" (MCP tool function name)
# - Actual execution may include more LLM calls (thinking, planning, etc.), which is normal
# ============================================================================
reference_trajectory:
# ===== Level 1: SPAN =====
- "SPAN: markdown_validation"
# ===== Level 2: Chain (wildcard handles UUID) =====
- "Chain: Crew***.kickoff"
# ===== Task: syntax_review_task =====
# Requirements Manager validates Markdown file and generates report
- "Agent: Requirements Manager"
- "LLM: *" # First call: decide to use MCP tool
- "Tool: markdown_validation_tool" # Call MCP validation tool
- "LLM: *" # Second call: generate JSON report
# Target tools list (for single-tool use metric)
# Only includes actual tool calls, used to detect tool usage
target_tools:
- "Tool: markdown_validation_tool"
# Models to evaluate
models:
- "GPT-5"
- "GPT-4o-mini"
- "DeepSeek-V3-1"
- "DeepSeek-R1"
- "Gemini-2.5-flash"
- "Gemini-2.5-flash-nothinking"
- "Qwen3-235b"
# ============================================================================
# Usage Instructions
# ============================================================================
#
# 1. Trajectory Hierarchy:
# - Level 1: SPAN (markdown_validation)
# - Level 2: Chain (Crew***.kickoff, wildcard handles UUID)
# - Level 3: AGENT (Requirements Manager, exact match)
# - Level 4: LLM (wildcard matches any model)
# - Level 4: Tool (markdown_validation_tool, exact match)
#
# 2. Wildcard Handling:
# - Crew_<UUID>.kickoff -> Crew***.kickoff (evaluate_trajectory.py auto-handles)
# - LLM: * -> Matches any model name (e.g. gpt-4o-mini, deepseek-r1, etc.)
#
# 3. Exact Match Requirements:
# - SPAN name: "markdown_validation" (exact match)
# Note: Evaluation script auto-filters nested second-level markdown_validation SPAN
# - Chain name: "Crew***.kickoff" (wildcard match)
# - AGENT name: "Requirements Manager" (exact match to role in agent.yaml)
# - LLM name: "LLM: *" (wildcard matches any model)
# - Tool name: "markdown_validation_tool" (exact match to MCP tool function name)
#
# 4. Evaluation Metrics Meaning:
# - Exact Match: Requires trajectory to be exactly the same (including LLM call count)
# - In-order Match: Allows extra calls, but core steps must appear in order
# - Any-order Match: Just needs to contain all necessary steps (ignores order)
# - Precision: Proportion of correct steps in predicted trajectory
# - Recall: Proportion of reference trajectory covered
# - Single-tool Use: Detects markdown_validation_tool usage
# - unique_path_ratio: Path diversity (unique complete trajectories / samples)
# - path_entropy: Path entropy (Shannon entropy based on trajectory frequency, normalized 0-1)
#
# 5. Run Command:
# cd /Users/wzr/TOSEM-2025/RESULTS/RQ-Failure_Breakdown/MarkdownValidator-MCP
# python3 evaluate_trajectory.py --config reference_trajectory.yaml
#
# 6. Design Notes:
# - This reference trajectory represents ideal execution path
# - Single markdown_validation SPAN (nested ones auto-filtered)
# - Requirements Manager: 2 LLM + 1 tool
# * First LLM: Analyze task and decide to use tool
# * Tool call: Validate Markdown file via MCP
# * Second LLM: Process results and generate JSON output
# - Total: 1 SPAN + 1 Chain + 1 Agent + 2 LLM + 1 tool
# - In-order Match and Any-order Match metrics better for evaluating actual performance
#
# 7. Differences from Actual Trajectories:
# - Actual trajectories may include more LLM calls (thinking, planning, execution, summary, etc.)
# - Some models may call validation tool multiple times (retry or extra checks)
# - These differences don't represent errors, but different model execution strategies
# - Exact Match score may be low, focus on In-order Match and Recall
#
# 8. MCP Version Characteristics:
# - Tools provided through MCP server, using SSE protocol communication
# - Tools obtained from server via MCPServerAdapter
# - Actual execution may have nested markdown_validation SPANs (main.py lines 280-281)
# - Evaluation script auto-filters nested SPANs to ensure evaluation consistency
# - Tool name is function name (markdown_validation_tool)
# - Main difference from non-MCP version: tool provision method (reference trajectory structure same)