AINativeBench / data /processed /RQ1 /RecruitmentAssistant-H_A2A /reference_trajectory.yaml
王子睿
restructure + add files
8c10cf2
Raw
History Blame Contribute Delete
15.4 kB
# Full reference trajectory configuration - RecruitmentAssistant-H_A2A
#
# Version: ported from RecruitmentAssistant-A2A and adapted to the A2A_mix hybrid architecture
# Principle: keep business logic unchanged while adapting to the hybrid framework hierarchy
# (LangGraph + AutoGen + CrewAI).
#
# A2A_mix hybrid architecture notes:
# - Stage 1 (Job Analysis): LangGraph
# * Use [Chain] LangGraph rather than CrewAI's Crew***.kickoff
# * The AGENT name is simplified to "agent" rather than the full role name
# * Tool calls appear under the [Chain] tools node and may run in batched mode
# * LangGraph-specific Chain nodes: _should_continue and format_output
#
# - Stage 2 (Candidate Evaluation): CrewAI
# * Keep the standard [Chain] Crew***.kickoff structure
# * Use the full agent role name: "Senior Candidate Evaluator._execute_core"
# * Tool calls are directly under AGENT (no intermediate Chain)
#
# - Stage 3 (Interview Communication): AutoGen
# * AGENT pattern: create_agent + invoke_agent
# * Ignore create_agent (excluded from evaluation)
# * Focus on invoke_agent execution
# * Tool call prefix: "execute_tool"
#
# Dynamic reference-trajectory design:
# 1. Choose 3x/4x variants based on the number of unified_web_search calls (consistent with A2A)
# 2. Dynamic matching for LangGraph tools batching:
# - 3 searches: 1+1+1, 1+2, 2+1, or 3
# - 4 searches: 1+1+1+1, 2+1+1, 1+2+1, 1+1+2, 3+1, 1+3, or 4
# 3. Dynamic matching for AutoGen Interview LLM/Tool patterns:
# Rule: must start and end with LLM; between adjacent tools, there can be 0 or 1 LLM
# - With 2 tools, possible patterns:
# * LLM → Tool1 → Tool2 → LLM (0 LLM between tools)
# * LLM → Tool1 → LLM → Tool2 → LLM (1 LLM between tools)
# - Enumeration: for each adjacent tool pair, independently choose inserting 0 or 1 LLM
# * 2 tools have 1 gap → 2^1 = 2 patterns
# * 3 tools have 2 gaps → 2^2 = 4 patterns
# 4. Tool-order permutation optimization for Candidate Evaluation / Interview Communication:
# - candidate_profile_analyzer ↔ candidate_evaluator_pro (2! = 2)
# - comprehensive_interview_material_generator ↔ email_template_generator (2! = 2)
# - Total permutations: 2 × 2 = 4
# During evaluation, all combinations (LangGraph × AutoGen LLM × tool permutations) are enumerated,
# and the variant with the highest score among exact_match / in_order_match / any_order_match is used.
# Project name
project_name: "RecruitmentAssistant-H_A2A"
# Extract types configuration
extract_types:
- "SPAN" # multi-level SPAN structure
- "Chain" # Chain nodes for LangGraph / CrewAI
- "AGENT" # Agent nodes across frameworks (AutoGen extracts invoke_agent only)
- "LLM" # LLM calls
- "Tool" # Tool calls
# ============================================================================
# Reference Trajectory (Ground Truth) - A2A_mix Hybrid Architecture
# ============================================================================
#
# Graphical structure (ideal execution path, 3x version example):
# Graphical structure (ideal execution path, 3x version example):
#
# [SPAN] recruitment_orchestrator
# └─ [SPAN] crew_execution
# ├─ [SPAN] analyze_job (LangGraph)
# │ └─ [SPAN] a2a_call_job_analysis
# │ └─ [SPAN] job_analysis_server_execution
# │ └─ [Chain] LangGraph
# │ ├─ [AGENT] agent
# │ │ ├─ [LLM] *
# │ │ └─ [Chain] _should_continue
# │ ├─ [Chain] tools (batched execution with 3 searches; possible groupings: 1+1+1, 1+2, 2+1, 3)
# │ │ ├─ [Tool] unified_web_search (1st)
# │ │ ├─ [Tool] unified_web_search (2nd)
# │ │ └─ [Tool] unified_web_search (3rd)
# │ ├─ [AGENT] agent
# │ │ ├─ [LLM] *
# │ │ └─ [Chain] _should_continue
# │ └─ [Chain] format_output
# ├─ [SPAN] evaluate_candidates (CrewAI)
# │ └─ [SPAN] a2a_call_candidate_evaluation
# │ └─ [SPAN] candidate_evaluation_server_execution
# │ └─ [Chain] Crew***.kickoff
# │ └─ [AGENT] Senior Candidate Evaluator
# │ ├─ [LLM] *
# │ ├─ [Tool] candidate_profile_analyzer
# │ ├─ [LLM] *
# │ ├─ [Tool] candidate_evaluator_pro
# │ └─ [LLM] *
# └─ [SPAN] prepare_interviews (AutoGen)
# └─ [SPAN] a2a_call_interview_communication
# └─ [SPAN] interview_communication_server_execution
# └─ [AGENT] invoke_agent interview_coordinator
# ├─ [LLM] *
# ├─ [Tool] execute_tool comprehensive_interview_material_generator
# ├─ [Tool] execute_tool email_template_generator
# └─ [LLM] *
#
# Notes:
# - Under LangGraph's [Chain] tools node, there may be multiple batched calls.
# For example, 4 searches may be split into two batches: 2 + 2.
# The evaluation script dynamically generates all possible grouping variants.
# - AutoGen's create_agent nodes are excluded from the reference trajectory.
# ============================================================================
# Default reference trajectory (compatibility; kept as the 3x version)
reference_trajectory:
# ===== Top-level SPAN =====
- "SPAN: recruitment_orchestrator"
# ===== Second-level SPAN (includes 3 stages) =====
- "SPAN: crew_execution"
# ===== Stage 1: Job Analysis (LangGraph, 3 searches) =====
- "SPAN: analyze_job"
- "SPAN: a2a_call_job_analysis"
- "SPAN: job_analysis_server_execution"
- "Chain: LangGraph"
- "AGENT: agent"
- "LLM: *"
- "Chain: _should_continue"
- "Chain: tools"
- "Tool: unified_web_search"
- "Tool: unified_web_search"
- "Tool: unified_web_search"
- "AGENT: agent"
- "LLM: *"
- "Chain: _should_continue"
- "Chain: format_output"
# ===== Stage 2: Candidate Evaluation (CrewAI) =====
- "SPAN: evaluate_candidates"
- "SPAN: a2a_call_candidate_evaluation"
- "SPAN: candidate_evaluation_server_execution"
- "Chain: Crew***.kickoff"
- "AGENT: Senior Candidate Evaluator"
- "LLM: *"
- "Tool: candidate_profile_analyzer"
- "LLM: *"
- "Tool: candidate_evaluator_pro"
- "LLM: *"
# ===== Stage 3: Interview Communication (AutoGen) =====
- "SPAN: prepare_interviews"
- "SPAN: a2a_call_interview_communication"
- "SPAN: interview_communication_server_execution"
- "AGENT: invoke_agent interview_coordinator"
- "LLM: *"
- "Tool: execute_tool comprehensive_interview_material_generator"
- "Tool: execute_tool email_template_generator"
- "LLM: *"
# ============================================================================
# Dynamic reference trajectory - Version 1: 3 unified_web_search calls
# ============================================================================
# This version serves as the base template for LangGraph tools batching.
# During evaluation, variants are generated dynamically based on the sample's tools grouping pattern.
reference_trajectory_3x:
# ===== Top-level SPAN =====
- "SPAN: recruitment_orchestrator"
# ===== Second-level SPAN (includes 3 stages) =====
- "SPAN: crew_execution"
# ===== Stage 1: Job Analysis (LangGraph, 3-search base template) =====
- "SPAN: analyze_job"
- "SPAN: a2a_call_job_analysis"
- "SPAN: job_analysis_server_execution"
- "Chain: LangGraph"
# First pass: agent decision
- "AGENT: agent"
- "LLM: *"
- "Chain: _should_continue"
# Tool groups (3 calls; may be split into 1-3 Chain: tools blocks)
- "Chain: tools"
- "Tool: unified_web_search"
- "Tool: unified_web_search"
- "Tool: unified_web_search"
# Second pass: agent summary
- "AGENT: agent"
- "LLM: *"
- "Chain: _should_continue"
- "Chain: format_output"
# ===== Stage 2: Candidate Evaluation (CrewAI) =====
- "SPAN: evaluate_candidates"
- "SPAN: a2a_call_candidate_evaluation"
- "SPAN: candidate_evaluation_server_execution"
- "Chain: Crew***.kickoff"
- "AGENT: Senior Candidate Evaluator"
- "LLM: *"
- "Tool: candidate_profile_analyzer"
- "LLM: *"
- "Tool: candidate_evaluator_pro"
- "LLM: *"
# ===== Stage 3: Interview Communication (AutoGen) =====
- "SPAN: prepare_interviews"
- "SPAN: a2a_call_interview_communication"
- "SPAN: interview_communication_server_execution"
- "AGENT: invoke_agent interview_coordinator"
- "LLM: *"
- "Tool: execute_tool comprehensive_interview_material_generator"
- "Tool: execute_tool email_template_generator"
- "LLM: *"
# ============================================================================
# Dynamic reference trajectory - Version 2: 4 unified_web_search calls
# ============================================================================
reference_trajectory_4x:
# ===== Top-level SPAN =====
- "SPAN: recruitment_orchestrator"
# ===== Second-level SPAN (includes 3 stages) =====
- "SPAN: crew_execution"
# ===== Stage 1: Job Analysis (LangGraph, 4-search base template) =====
- "SPAN: analyze_job"
- "SPAN: a2a_call_job_analysis"
- "SPAN: job_analysis_server_execution"
- "Chain: LangGraph"
# First pass: agent decision
- "AGENT: agent"
- "LLM: *"
- "Chain: _should_continue"
# Tool groups (4 calls; may be split into 1-4 Chain: tools blocks)
- "Chain: tools"
- "Tool: unified_web_search"
- "Tool: unified_web_search"
- "Tool: unified_web_search"
- "Tool: unified_web_search"
# Second pass: agent summary
- "AGENT: agent"
- "LLM: *"
- "Chain: _should_continue"
- "Chain: format_output"
# ===== Stage 2: Candidate Evaluation (CrewAI) =====
- "SPAN: evaluate_candidates"
- "SPAN: a2a_call_candidate_evaluation"
- "SPAN: candidate_evaluation_server_execution"
- "Chain: Crew***.kickoff"
- "AGENT: Senior Candidate Evaluator"
- "LLM: *"
- "Tool: candidate_profile_analyzer"
- "LLM: *"
- "Tool: candidate_evaluator_pro"
- "LLM: *"
# ===== Stage 3: Interview Communication (AutoGen) =====
- "SPAN: prepare_interviews"
- "SPAN: a2a_call_interview_communication"
- "SPAN: interview_communication_server_execution"
- "AGENT: invoke_agent interview_coordinator"
- "LLM: *"
- "Tool: execute_tool comprehensive_interview_material_generator"
- "Tool: execute_tool email_template_generator"
- "LLM: *"
# Target tool list (for the single-tool use metric)
target_tools:
- "Tool: unified_web_search"
- "Tool: candidate_profile_analyzer"
- "Tool: candidate_evaluator_pro"
- "Tool: execute_tool comprehensive_interview_material_generator"
- "Tool: execute_tool email_template_generator"
# Model list to evaluate
models:
- "GPT-5"
- "GPT-4o-mini"
- "DeepSeek-V3-1"
- "DeepSeek-R1"
- "Gemini-2.5-flash"
- "Gemini-2.5-flash-nothinking"
- "Qwen3-235b"
# Permutable tool groups (for dynamic tool-order optimization)
permutable_tool_groups:
# Candidate Evaluation stage tools can be in any order
candidate_evaluation_tools:
- "Tool: candidate_profile_analyzer"
- "Tool: candidate_evaluator_pro"
# Interview Communication stage tools can be in any order
interview_communication_tools:
- "Tool: execute_tool comprehensive_interview_material_generator"
- "Tool: execute_tool email_template_generator"
# ============================================================================
# Usage notes - A2A_mix hybrid architecture version
# ============================================================================
#
# 1. Hybrid framework notes:
# - Stage 1 (Job Analysis): LangGraph
# * Chain name: LangGraph (not Crew***.kickoff)
# * Simplified agent name: agent (not the full role name)
# * Batched tool execution: [Chain] tools may contain multiple Tool nodes
# * Special chain nodes: _should_continue and format_output
#
# - Stage 2 (Candidate Evaluation): CrewAI
# * Standard CrewAI structure: Crew***.kickoff → AGENT → LLM/Tool
# * Full agent name: Senior Candidate Evaluator
#
# - Stage 3 (Interview Communication): AutoGen
# * Ignore create_agent (excluded from evaluation)
# * Focus on invoke_agent interview_coordinator
# * Tool prefix: execute_tool
# 2. LangGraph tools batching dynamic mode:
# The evaluation script detects the tools grouping pattern in the sample and
# dynamically generates corresponding reference variants for matching.
#
# Possible groupings for 3 searches:
# - [1,1,1]: 3 Chain: tools blocks, each with 1 Tool
# - [1,2]: 2 blocks: first has 1 Tool, second has 2 Tools
# - [2,1]: 2 blocks: first has 2 Tools, second has 1 Tool
# - [3]: 1 block with 3 Tools
#
# Possible groupings for 4 searches:
# - [1,1,1,1]: 4 blocks, each with 1 Tool
# - [2,1,1]: 3 blocks with 2,1,1 Tools
# - [1,2,1]: 3 blocks with 1,2,1 Tools
# - [1,1,2]: 3 blocks with 1,1,2 Tools
# - [3,1]: 2 blocks with 3,1 Tools
# - [1,3]: 2 blocks with 1,3 Tools
# - [4]: 1 block with 4 Tools
# 3. AutoGen Interview LLM/Tool dynamic mode:
# The evaluation script detects the LLM/Tool call pattern in the AutoGen part and
# dynamically generates corresponding reference variants for matching.
#
# Supported patterns:
# - Compact: LLM → Tool1 → Tool2 → LLM (no LLM between tools)
# - Interleaved: LLM → Tool1 → LLM → Tool2 → LLM (insert one LLM between tools)
#
# Validation rules:
# - Must start and end with LLM
# - Must include the 2 target tools in the middle (fixed order)
# - At most one LLM between tools
# 4. Dynamic reference selection strategy (enhanced):
# For each sample:
# a) Choose the 3x/4x base template by unified_web_search total count
# b) Detect the LangGraph tools actual grouping pattern
# c) Detect the AutoGen Interview actual LLM/Tool pattern
# d) Enumerate all combinations (LangGraph pattern × AutoGen pattern)
# - Example: 3 searches have 4 LangGraph patterns and 2 AutoGen patterns
# - Total: 4 × 2 = 8 reference variants
# e) Compute exact_match, in_order_match, and any_order_match for each variant
# f) Select the variant with the best score as the final reference for the sample
# 5. Metric definitions (consistent with A2A):
# - Exact Match: prediction equals reference
# - In-order Match: reference is a subsequence of prediction
# - Any-order Match: contains all required steps (order ignored)
# - Precision: fraction of correct steps in prediction
# - Recall: fraction of reference steps covered
# - Single-tool Use: average usage rate of target tools
# - unique_path_ratio: path diversity
# - path_entropy: path entropy
# 6. Command:
# cd /Users/wzr/TOSEM-2025/RESULTS/RQ-Failure_Breakdown/RecruitmentAssistant-H_A2A
# python3 evaluate_trajectory.py --config reference_trajectory.yaml
# 7. Key differences vs the A2A version:
# - LangGraph-specific Chain node structure
# - Simplified agent name (LangGraph uses "agent")
# - AutoGen invoke_agent pattern
# - Dynamic handling of LangGraph tools batching
# - Dynamic handling of AutoGen LLM/Tool patterns (new)
# - AutoGen tool prefix: execute_tool (instead of CrewAI's ._use)
# - Dual dynamic matching: supports LangGraph and AutoGen pattern combinations