AINativeBench / data /processed /RQ1 /LandingPageGenerator-MCP /reference_trajectory.yaml
王子睿
restructure + add files
8c10cf2
Raw
History Blame Contribute Delete
9.67 kB
# Full reference trajectory configuration - LandingPageGenerator-MCP project
#
# Version: ideal trajectory based on code design (applies to all models)
# Principle: only include steps explicitly required by the code design; do not rely on test statistics
#
# Code design analysis:
# 1. Three Crews run sequentially (orchestrator.py):
# - IdeaExpansionCrew: Product Idea Analyst uses bocha_websearch_tool for market research
# - TemplateSelectionCrew: Template Selection Specialist uses learn_landing_page_options to select templates
# - ContentCreationCrew: Landing Page Content Generator uses read_file_content and write_file_with_content to generate HTML
#
# 2. Tool usage analysis:
# - Crew 1: bocha_websearch_tool (MCP tool; web search)
# - Crew 2: learn_landing_page_options (MCP tool; learn available templates)
# - Crew 3: read_file_content + write_file_with_content (MCP tools; file operations)
#
# 3. Execution flow:
# - orchestrator creates the top-level SPAN: landing_page_generation-MCP
# - orchestrator creates crew_execution SPAN containing 3 child SPANs
# - Crew SPANs: idea_expansion, template_selection, content_creation
# - Each Crew includes one Chain and one Agent
# - The Agent uses the LLM to decide tool calls; after tool returns, the LLM generates the output
#
# 4. SPAN hierarchy:
# - landing_page_generation-MCP (top-level; orchestrator.py line 341)
# └─ crew_execution (contains 3 Crews; orchestrator.py line 365)
# ├─ idea_expansion (Crew 1; orchestrator.py line 70)
# ├─ template_selection (Crew 2; orchestrator.py line 115)
# └─ content_creation (Crew 3; orchestrator.py line 167)
# Project name
project_name: "LandingPageGenerator-MCP"
# Trajectory extraction type configuration
extract_types:
- "SPAN" # Multi-level SPAN structure
- "Chain" # Kickoff chain for each Crew
- "AGENT" # Agent for each Crew
- "LLM" # LLM calls
- "Tool" # MCP tool calls
# ============================================================================
# Reference trajectory (Ground Truth) - based on code design
# ============================================================================
#
# Diagram (ideal execution path):
#
# [SPAN] landing_page_generation-MCP
# └─ [SPAN] crew_execution
# ├─ [SPAN] idea_expansion
# │ └─ [Chain] Crew***.kickoff
# │ └─ [AGENT] Product Idea Analyst
# │ ├─ [LLM] *
# │ ├─ [Tool] bocha_websearch_tool
# │ └─ [LLM] *
# ├─ [SPAN] template_selection
# │ └─ [Chain] Crew***.kickoff
# │ └─ [AGENT] Template Selection Specialist
# │ ├─ [LLM] *
# │ ├─ [Tool] learn_landing_page_options
# │ └─ [LLM] *
# └─ [SPAN] content_creation
# └─ [Chain] Crew***.kickoff
# └─ [AGENT] Landing Page Content Generator
# ├─ [LLM] *
# ├─ [Tool] read_file_content
# ├─ [LLM] *
# ├─ [Tool] write_file_with_content
# └─ [LLM] *
#
# Design basis:
# - orchestrator.py: runs 3 Crews sequentially
# - Each Crew has 1 Agent and 1 Task
# - idea_expansion_crew.py: Product Idea Analyst uses bocha_websearch_tool
# - template_selection_crew.py: Template Selection Specialist uses learn_landing_page_options
# - content_creation_crew.py: Landing Page Content Generator uses read_file_content and write_file_with_content
# - All Agent roles are defined in agents.yaml
# - Tools are retrieved from the MCP server via MCPServerAdapter
#
# Ideal trajectory notes:
# - Top-level SPAN: landing_page_generation-MCP
# - Second-level SPAN: crew_execution (contains 3 Crews)
# - The 3 Crews run sequentially; each has its own SPAN
# - Each Agent: at least 2 LLM calls + required tool calls
# - Idea Expansion: 1 tool call (bocha_websearch_tool)
# - Template Selection: 1 tool call (learn_landing_page_options)
# - Content Creation: 2 tool calls (read_file_content + write_file_with_content)
# - This trajectory represents the ideal execution path with no retries and no redundant steps
#
# Notes:
# - "LLM: *" means any LLM model (wildcard match)
# - Agent names must exactly match the role fields in agents.yaml
# - Tool names must exactly match the tool names provided by the MCP server
# - Actual execution may include more LLM calls (thinking/planning/etc.); this is expected
# ============================================================================
reference_trajectory:
# ===== Top-level SPAN =====
- "SPAN: landing_page_generation-MCP"
# ===== Second-level SPAN (contains 3 Crews) =====
- "SPAN: crew_execution"
# ===== Crew 1: Idea Expansion =====
- "SPAN: idea_expansion"
- "Chain: Crew***.kickoff"
- "Agent: Product Idea Analyst"
- "LLM: *" # Decide to call tools
- "Tool: bocha_websearch_tool" # Web search
- "LLM: *" # Generate an expanded idea
# ===== Crew 2: Template Selection =====
- "SPAN: template_selection"
- "Chain: Crew***.kickoff"
- "Agent: Template Selection Specialist"
- "LLM: *" # Decide to call tools
- "Tool: learn_landing_page_options" # Learn available templates
- "LLM: *" # Select a template
# ===== Crew 3: Content Creation =====
- "SPAN: content_creation"
- "Chain: Crew***.kickoff"
- "Agent: Landing Page Content Generator"
- "LLM: *" # Decide to read the template
- "Tool: read_file_content" # Read the template file
- "LLM: *" # Generate HTML content
- "Tool: write_file_with_content" # Write the HTML file
- "LLM: *" # Final confirmation
# Target tool list (used by the single-tool use metric)
# List all required tools here
target_tools:
- "Tool: bocha_websearch_tool"
- "Tool: learn_landing_page_options"
- "Tool: read_file_content"
- "Tool: write_file_with_content"
# 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
# ============================================================================
#
# 1. Trajectory hierarchy:
# - Level 1: SPAN (landing_page_generation-MCP)
# - Level 2: SPAN (idea_expansion, template_selection, content_creation)
# - Level 3: Chain (Crew***.kickoff; wildcard-normalized UUID)
# - Level 4: AGENT (exact match on agent role)
# - Level 5: LLM (wildcard match any model)
# - Level 5: Tool (exact match on tool name)
#
# 2. Wildcard handling:
# - Crew_<UUID>.kickoff -> Crew***.kickoff (handled automatically by evaluate_trajectory.py)
# - LLM: * -> match any model name (e.g., gpt-4o-mini, deepseek-r1)
#
# 3. Exact match requirements:
# - SPAN names: "landing_page_generation-MCP", "crew_execution", "idea_expansion", "template_selection", "content_creation"
# - Chain names: "Crew***.kickoff" (wildcard match)
# - AGENT names: "Product Idea Analyst", "Template Selection Specialist", "Landing Page Content Generator"
# (exact match the role fields in agents.yaml)
# - LLM: "LLM: *" (wildcard match any model)
# - Tool: exact match MCP tool names (bocha_websearch_tool, learn_landing_page_options,
# read_file_content, write_file_with_content)
#
# 4. Metric meanings:
# - Exact Match: trajectories must be identical (including LLM call count)
# - In-order Match: extra calls allowed, but core steps must appear in order
# - Any-order Match: contains all required steps (order ignored)
# - Precision: fraction of predicted steps considered correct
# - Recall: fraction of reference steps covered
# - Single-tool Use: checks usage of all 4 required tools
# - unique_path_ratio: path diversity (unique full trajectories / sample count)
# - path_entropy: path entropy (Shannon entropy over trajectory frequencies; normalized to 0-1)
#
# 5. Run command:
# cd /Users/wzr/TOSEM-2025/RESULTS/RQ-Failure_Breakdown/LandingPageGenerator-MCP
# python3 evaluate_trajectory.py --config reference_trajectory.yaml
#
# 6. Design notes:
# - This reference trajectory represents the ideal execution path
# - Three Crews run sequentially; each Crew completes its task independently
# - Crew 1 (Idea Expansion): 1 LLM decision + 1 web search + 1 LLM generation
# - Crew 2 (Template Selection): 1 LLM decision + 1 template-learning tool call + 1 LLM selection
# - Crew 3 (Content Creation): 1 LLM decision + 1 template read + 1 LLM generation + 1 file write + 1 LLM confirmation
# - Total: 2 top-level SPANs (landing_page_generation-MCP + crew_execution) + 3 Crew SPANs + 3 Chains + 3 Agents + 9 LLM calls + 4 tool calls
# - In-order Match and Recall are often more suitable for evaluating real performance
#
# 7. Differences vs. actual trajectories:
# - Actual trajectories may include more LLM calls (thinking/planning/execution/summarization)
# - Some models may call tools multiple times (retries or extra checks)
# - Template Selection may call learn_landing_page_options multiple times
# - These differences are not necessarily errors; they reflect different execution strategies
# - Exact Match may be low; focus on In-order Match and Recall
#
# 8. MCP version characteristics:
# - All tools are provided via the MCP server using SSE for communication
# - Tools are fetched from the server via MCPServerAdapter
# - Includes a multi-level SPAN structure (orchestrator + crew_execution + 3 Crews)
# - Three independent Crews collaborate to complete the landing page generation task
# - Each Crew has clear responsibilities and required tools