AINativeBench / data /processed /RQ1 /RecruitmentAssistant-MCP /reference_trajectory.yaml
王子睿
restructure + add files
8c10cf2
Raw
History Blame Contribute Delete
19.1 kB
# Full reference trajectory configuration - RecruitmentAssistant-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):
# - JobAnalysisCrew: Senior Job Requirements Analyst uses unified_web_search for market research
# - CandidateEvaluationCrew: Senior Candidate Evaluator uses candidate_evaluator_pro to evaluate candidates
# - InterviewCommunicationCrew: Senior Interview Materials Coordinator uses tools to generate interview materials and email templates
#
# 2. Tool usage analysis:
# - Crew 1: unified_web_search (MCP tool; web search for market research)
# - Crew 2: candidate_evaluator_pro (MCP tool; candidate evaluation and scoring)
# - Crew 3: comprehensive_interview_material_generator (MCP tool; interview materials generation) +
# email_template_generator (MCP tool; email template generation)
#
# 3. Execution flow:
# - orchestrator creates the top-level SPAN: recruitment_orchestrator
# - orchestrator creates crew_execution SPAN containing 3 child SPANs
# - Crew SPANs: analyze_job, evaluate_candidates, prepare_interviews
# - 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:
# - recruitment_orchestrator (top-level; orchestrator.py line 353)
# └─ crew_execution (contains 3 Crews; orchestrator.py line 371)
# ├─ analyze_job (Crew 1; orchestrator.py line 141)
# ├─ evaluate_candidates (Crew 2; orchestrator.py line 190)
# └─ prepare_interviews (Crew 3; orchestrator.py line 238)
# Project name
project_name: "RecruitmentAssistant-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):
#
# Note: dynamically select the reference trajectory based on the actual number of unified_web_search calls:
# - If unified_web_search under analyze_job SPAN <= 3, use reference_trajectory_3x
# - If unified_web_search under analyze_job SPAN >= 4, use reference_trajectory_4x
#
# === Version 1: 3x unified_web_search ===
# [SPAN] recruitment_orchestrator
# └─ [SPAN] crew_execution
# ├─ [SPAN] analyze_job
# │ └─ [Chain] Crew***.kickoff
# │ └─ [AGENT] Senior Job Requirements Analyst
# │ ├─ [LLM] *
# │ ├─ [Tool] unified_web_search (1st search)
# │ ├─ [LLM] *
# │ ├─ [Tool] unified_web_search (2nd search)
# │ ├─ [LLM] *
# │ ├─ [Tool] unified_web_search (3rd search)
# │ └─ [LLM] *
# ├─ [SPAN] evaluate_candidates
# │ └─ [Chain] Crew***.kickoff
# │ └─ [AGENT] Senior Candidate Evaluator
# │ ├─ [LLM] *
# │ ├─ [Tool] candidate_evaluator_pro
# │ └─ [LLM] *
# └─ [SPAN] prepare_interviews
# └─ [Chain] Crew***.kickoff
# └─ [AGENT] Senior Interview Materials Coordinator
# ├─ [LLM] *
# ├─ [Tool] comprehensive_interview_material_generator
# ├─ [LLM] *
# ├─ [Tool] email_template_generator
# └─ [LLM] *
#
# === Version 2: 4x unified_web_search ===
# [SPAN] recruitment_orchestrator
# └─ [SPAN] crew_execution
# ├─ [SPAN] analyze_job
# │ └─ [Chain] Crew***.kickoff
# │ └─ [AGENT] Senior Job Requirements Analyst
# │ ├─ [LLM] *
# │ ├─ [Tool] unified_web_search (1st search)
# │ ├─ [LLM] *
# │ ├─ [Tool] unified_web_search (2nd search)
# │ ├─ [LLM] *
# │ ├─ [Tool] unified_web_search (3rd search)
# │ ├─ [LLM] *
# │ ├─ [Tool] unified_web_search (4th search)
# │ └─ [LLM] *
# ├─ [SPAN] evaluate_candidates
# │ └─ [Chain] Crew***.kickoff
# │ └─ [AGENT] Senior Candidate Evaluator
# │ ├─ [LLM] *
# │ ├─ [Tool] candidate_evaluator_pro
# │ └─ [LLM] *
# └─ [SPAN] prepare_interviews
# └─ [Chain] Crew***.kickoff
# └─ [AGENT] Senior Interview Materials Coordinator
# ├─ [LLM] *
# ├─ [Tool] comprehensive_interview_material_generator
# ├─ [LLM] *
# ├─ [Tool] email_template_generator
# └─ [LLM] *
#
# Design basis:
# - orchestrator.py: runs 3 Crews sequentially, with retry support
# - Each Crew has 1 Agent and 1 Task
# - job_analysis_crew.py: Senior Job Requirements Analyst uses unified_web_search
# * agents.yaml: requires 3-5 targeted searches
# * tasks.yaml: requires searching market data, salary standards, industry trends, etc.
# - candidate_evaluation_crew.py: Senior Candidate Evaluator uses candidate_evaluator_pro
# * agents.yaml: uses candidate_evaluator_pro to score, rank, and grade all candidates
# * tasks.yaml: must use the candidate_evaluator_pro tool
# - interview_communication_crew.py: Senior Interview Materials Coordinator uses 2 tools
# * agents.yaml: comprehensive_interview_material_generator (interview_stage="all")
# email_template_generator (template_type="all")
# * tasks.yaml: explicitly requires calling these 2 tools
# - All Agent roles are defined in their respective agents.yaml
# - Tools are retrieved from the MCP server via MCPServerAdapter
#
# Ideal trajectory notes:
# - Top-level SPAN: recruitment_orchestrator
# - 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
# - Job Analysis: multiple tool calls (unified_web_search; 3-4x for market research)
# - Candidate Evaluation: 1 tool call (candidate_evaluator_pro)
# - Interview Communication: 2 tool calls (comprehensive_interview_material_generator + email_template_generator)
# - 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
# - unified_web_search may be called multiple times (3-4x), which is normal for market research
# - Actual execution may include more LLM calls (thinking/planning/etc.); this is expected
# ============================================================================
# [Deprecated] Original reference trajectory (kept for compatibility)
# Now dynamically select between reference_trajectory_3x and reference_trajectory_4x
reference_trajectory:
# Default to the 3x version (backward compatible)
- "SPAN: recruitment_orchestrator"
- "SPAN: crew_execution"
- "SPAN: analyze_job"
- "Chain: Crew***.kickoff"
- "Agent: Senior Job Requirements Analyst"
- "LLM: *"
- "Tool: unified_web_search"
- "LLM: *"
- "Tool: unified_web_search"
- "LLM: *"
- "Tool: unified_web_search"
- "LLM: *"
- "SPAN: evaluate_candidates"
- "Chain: Crew***.kickoff"
- "Agent: Senior Candidate Evaluator"
- "LLM: *"
- "Tool: candidate_evaluator_pro"
- "LLM: *"
- "SPAN: prepare_interviews"
- "Chain: Crew***.kickoff"
- "Agent: Senior Interview Materials Coordinator"
- "LLM: *"
- "Tool: comprehensive_interview_material_generator"
- "LLM: *"
- "Tool: email_template_generator"
- "LLM: *"
# ============================================================================
# Dynamic reference trajectory - Version 1: 3x unified_web_search
# ============================================================================
# Applies when unified_web_search count under analyze_job SPAN <= 3
reference_trajectory_3x:
# ===== Top-level SPAN =====
- "SPAN: recruitment_orchestrator"
# ===== Second-level SPAN (contains 3 Crews) =====
- "SPAN: crew_execution"
# ===== Crew 1: Job Analysis (3 searches) =====
- "SPAN: analyze_job"
- "Chain: Crew***.kickoff"
- "Agent: Senior Job Requirements Analyst"
- "LLM: *" # Decide to call tools
- "Tool: unified_web_search" # Market research - search #1
- "LLM: *" # Analyze results and decide to continue searching
- "Tool: unified_web_search" # Market research - search #2
- "LLM: *" # Analyze results and decide to continue searching
- "Tool: unified_web_search" # Market research - search #3
- "LLM: *" # Produce the final job analysis report
# ===== Crew 2: Candidate Evaluation =====
- "SPAN: evaluate_candidates"
- "Chain: Crew***.kickoff"
- "Agent: Senior Candidate Evaluator"
- "LLM: *" # Decide to call tools
- "Tool: candidate_profile_analyzer" # Analyze candidate profiles (optional; recommended)
- "LLM: *" # Process analysis results
- "Tool: candidate_evaluator_pro" # Evaluate all candidates (scoring, ranking, grading)
- "LLM: *" # Produce the final evaluation report
# ===== Crew 3: Interview Communication =====
- "SPAN: prepare_interviews"
- "Chain: Crew***.kickoff"
- "Agent: Senior Interview Materials Coordinator"
- "LLM: *" # Decide to generate interview materials
- "Tool: comprehensive_interview_material_generator" # Generate interview questions for all stages
- "LLM: *" # Review materials and decide to generate email templates
- "Tool: email_template_generator" # Generate all types of email templates
- "LLM: *" # Final merge to produce a complete interview package
# ============================================================================
# Dynamic reference trajectory - Version 2: 4x unified_web_search
# ============================================================================
# Applies when unified_web_search count under analyze_job SPAN >= 4
reference_trajectory_4x:
# ===== Top-level SPAN =====
- "SPAN: recruitment_orchestrator"
# ===== Second-level SPAN (contains 3 Crews) =====
- "SPAN: crew_execution"
# ===== Crew 1: Job Analysis (4 searches) =====
- "SPAN: analyze_job"
- "Chain: Crew***.kickoff"
- "Agent: Senior Job Requirements Analyst"
- "LLM: *" # Decide to call tools
- "Tool: unified_web_search" # Market research - search #1
- "LLM: *" # Analyze results and decide to continue searching
- "Tool: unified_web_search" # Market research - search #2
- "LLM: *" # Analyze results and decide to continue searching
- "Tool: unified_web_search" # Market research - search #3
- "LLM: *" # Analyze results and decide to continue searching
- "Tool: unified_web_search" # Market research - search #4
- "LLM: *" # Produce the final job analysis report
# ===== Crew 2: Candidate Evaluation =====
- "SPAN: evaluate_candidates"
- "Chain: Crew***.kickoff"
- "Agent: Senior Candidate Evaluator"
- "LLM: *" # Decide to call tools
- "Tool: candidate_profile_analyzer" # Analyze candidate profiles (optional; recommended)
- "LLM: *" # Process analysis results
- "Tool: candidate_evaluator_pro" # Evaluate all candidates (scoring, ranking, grading)
- "LLM: *" # Produce the final evaluation report
# ===== Crew 3: Interview Communication =====
- "SPAN: prepare_interviews"
- "Chain: Crew***.kickoff"
- "Agent: Senior Interview Materials Coordinator"
- "LLM: *" # Decide to generate interview materials
- "Tool: comprehensive_interview_material_generator" # Generate interview questions for all stages
- "LLM: *" # Review materials and decide to generate email templates
- "Tool: email_template_generator" # Generate all types of email templates
- "LLM: *" # Final merge to produce a complete interview package
# Target tool list (used by the single-tool use metric)
# List all required tools here
target_tools:
- "Tool: unified_web_search"
- "Tool: candidate_profile_analyzer"
- "Tool: candidate_evaluator_pro"
- "Tool: comprehensive_interview_material_generator"
- "Tool: email_template_generator"
# ============================================================================
# Dynamic tool permutation configuration
# ============================================================================
# Define permutable tool groups for certain stages to generate reference trajectories in different orders
# During evaluation, all permutations are tried and the best-matching one is selected as the reference
permutable_tool_groups:
# Tools in the Candidate Evaluation stage can be in any order
candidate_evaluation_tools:
- "Tool: candidate_profile_analyzer"
- "Tool: candidate_evaluator_pro"
# Tools in the Interview Communication stage can be in any order
interview_communication_tools:
- "Tool: comprehensive_interview_material_generator"
- "Tool: email_template_generator"
# 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 (recruitment_orchestrator)
# - Level 2: SPAN (crew_execution)
# - Level 3: SPAN (analyze_job, evaluate_candidates, prepare_interviews)
# - Level 4: Chain (Crew***.kickoff; wildcard-normalized UUID)
# - Level 5: AGENT (exact match on agent role)
# - Level 6: LLM (wildcard match any model)
# - Level 6: 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-5-chat-latest, deepseek-reasoner)
#
# 3. Exact match requirements:
# - SPAN names: "recruitment_orchestrator", "crew_execution", "analyze_job",
# "evaluate_candidates", "prepare_interviews"
# - Chain names: "Crew***.kickoff" (wildcard match)
# - AGENT names: "Senior Job Requirements Analyst", "Senior Candidate Evaluator",
# "Senior Interview Materials Coordinator"
# (exact match the role fields in agents.yaml)
# - LLM: "LLM: *" (wildcard match any model)
# - Tool: exact match MCP tool names (unified_web_search, candidate_evaluator_pro,
# comprehensive_interview_material_generator, email_template_generator)
#
# 4. Metric meanings:
# - Exact Match: trajectories must be identical (including the counts of LLM/tool calls)
# - 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/RecruitmentAssistant-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 (Job Analysis):
# * 3x: multiple LLM calls + 3 web searches + LLM report generation
# * 4x: multiple LLM calls + 4 web searches + LLM report generation
# - Crew 2 (Candidate Evaluation): LLM decision + 1 candidate evaluation tool call + LLM report
# - Crew 3 (Interview Communication): LLM decision + 1 interview-material generation + LLM review + 1 email-template generation + LLM merge
# - 3x total: 2 top-level SPANs + 3 Crew SPANs + 3 Chains + 3 Agents + multiple LLM calls + 6 tool calls
# - 4x total: 2 top-level SPANs + 3 Crew SPANs + 3 Chains + 3 Agents + multiple LLM calls + 7 tool calls
# - In-order Match and Recall are often more suitable for evaluating real performance
#
# 7. Dynamic reference trajectory selection:
# - The evaluation script automatically counts unified_web_search calls under analyze_job SPAN for each sample
# - If count <= 3, use reference_trajectory_3x as the reference
# - If count >= 4, use reference_trajectory_4x as the reference
# - This dynamic selection makes evaluation fairer and adapts to different model search strategies
#
# 8. Differences vs. actual trajectories:
# - Actual trajectories may include more LLM calls (thinking/planning/etc.)
# - Job Analysis may call unified_web_search more times (3-5 is normal)
# - Some models may not use candidate_evaluator_pro and instead evaluate manually
# - Some models may call email_template_generator 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
#
# 9. 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 recruitment workflow
# - Each Crew has clear responsibilities and required tools
# - Supports orchestrator-level retry (max_crew_retries)
#
# 10. Tool usage notes:
# - unified_web_search: for market research (salary standards, industry trends, similar roles, etc.)
# * Recommended 3-4 calls, each targeting a different research angle
# * Params: query (search query), count (number of results; default 5), search_type (type)
# - candidate_evaluator_pro: comprehensive evaluation for all candidates
# * Includes technical score, experience score, cultural fit score
# * Auto ranking and grading (A/B/C/D)
# * Params: candidates_data (candidate data), job_requirements (job requirements)
# - comprehensive_interview_material_generator: generate a complete interview question bank
# * Use interview_stage="all" to generate all stages (screening/technical/behavioral/final)
# * Params: candidate_name, job_title, interview_stage, candidate_background, special_focus_areas
# - email_template_generator: generate various email templates
# * Use template_type="all" to generate all templates (outreach/interview_invite/offer/rejection/follow_up)
# * Params: template_type, candidate_name, job_title, company_name, custom_details