File size: 14,046 Bytes
8c10cf2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 | # Full reference-trajectory configuration - SocialMediaManager-H_A2A
#
# Version: ported from SocialMediaManager-A2A and adapted to the A2A_mix hybrid architecture
# Principle: keep business logic unchanged while adapting to the hybrid framework hierarchy (LangGraph + CrewAI + AutoGen)
#
# A2A_mix hybrid-architecture characteristics:
# - Stage 1 (Topic Analysis): LangGraph
# * Uses [Chain] LangGraph instead of CrewAI's Crew***.kickoff
# * Uses a simplified AGENT name: "agent" (not the full role name)
# * Tool calls appear under [Chain] tools, supporting batched execution
# * LangGraph-specific chains: _should_continue and format_output
#
# - Stage 2 (Content Generation): CrewAI
# * Keeps the standard [Chain] Crew***.kickoff structure
# * Uses the full agent role name: "Shakespearean Bard._execute_core"
# * Tool calls appear directly under AGENT (no intermediate Chain)
#
# - Stage 3 (Post Review): AutoGen
# * AGENT pattern: create_agent + invoke_agent
# * Ignore create_agent (excluded from evaluation)
# * Focus on invoke_agent execution
# * Tool prefix is "execute_tool"
# * Filter all [SPAN] mcp client/operation and its children
#
# Dynamic reference-trajectory design (5-way dynamic composition):
# 1. LangGraph tools batched-execution grouping (2 options):
# - [1,1]: two Chain: tools, each containing 1 tool
# - [2]: one Chain: tools containing 2 tools
#
# 2. LangGraph tools order (2! = 2 options):
# - keyword_extractor → topic_complexity_analyzer
# - topic_complexity_analyzer → keyword_extractor
#
# 3. AutoGen LLM/Tool insertion pattern (2^4 = 16 options):
# Rule: must start and end with an LLM; between adjacent tools there may be 0 or 1 LLM
# - 5 tools have 4 adjacency gaps → 2^4 = 16 combinations
# - Example: LLM → Tool1 → Tool2 → Tool3 → Tool4 → Tool5 → LLM (no insertion)
# LLM → Tool1 → LLM → Tool2 → LLM → Tool3 → LLM → Tool4 → LLM → Tool5 → LLM (insert everywhere)
#
# 4. AutoGen tools order (5! = 120 options):
# - The 5 verification tools can appear in any order
# - character_counter_tool, emoji_detector_tool, post_structure_validator,
# shakespearean_style_detector, tone_analyzer
#
# 5. CrewAI tool order (5! = 120 options):
# - Shakespearean Bard's 5 validation tools can appear in any order
# - keyword_extractor must be the first tool (not permuted)
#
# Total combinations: 2 × 2 × 16 × 120 × 120 = 921,600 reference variants
#
# Evaluation strategy: for each sample, select the variant with the highest score:
# exact_match×3 + in_order_match×2 + any_order_match×1
# Project name
project_name: "SocialMediaManager-H_A2A"
# Extraction types
extract_types:
- "SPAN" # multi-level SPAN structure
- "Chain" # Chain nodes for LangGraph/CrewAI
- "AGENT" # Agent nodes (AutoGen: only invoke_agent is extracted)
- "LLM" # LLM calls
- "Tool" # tool calls
# ============================================================================
# Reference trajectory (Ground Truth) - A2A_mix hybrid architecture
# ============================================================================
#
# Visual structure (ideal execution path; example with batched tools execution):
#
# [SPAN] shakespeare_x_post_orchestrator
# └─ [SPAN] crew_execution
# ├─ [SPAN] analyze_topic (LangGraph)
# │ └─ [SPAN] a2a_call_topic_analyzer
# │ └─ [SPAN] topic_analyzer_server_execution
# │ └─ [Chain] LangGraph
# │ ├─ [AGENT] agent
# │ │ ├─ [LLM] *
# │ │ └─ [Chain] _should_continue
# │ ├─ [Chain] tools (batched execution of 2 tools; possible groupings: 1+1, 2)
# │ │ ├─ [Tool] keyword_extractor
# │ │ └─ [Tool] topic_complexity_analyzer
# │ ├─ [AGENT] agent
# │ │ ├─ [LLM] *
# │ │ └─ [Chain] _should_continue
# │ └─ [Chain] format_output
# └─ [SPAN] content_generation_loop
# ├─ [SPAN] a2a_call_content_generator (CrewAI)
# │ └─ [SPAN] content_generator_server_execution
# │ └─ [Chain] Crew***.kickoff
# │ └─ [AGENT] Shakespearean Bard
# │ ├─ [LLM] *
# │ ├─ [Tool] keyword_extractor
# │ ├─ [LLM] *
# │ ├─ [Tool] character_counter_tool
# │ ├─ [LLM] *
# │ ├─ [Tool] emoji_detector_tool
# │ ├─ [LLM] *
# │ ├─ [Tool] post_structure_validator
# │ ├─ [LLM] *
# │ ├─ [Tool] shakespearean_style_detector
# │ ├─ [LLM] *
# │ ├─ [Tool] tone_analyzer
# │ └─ [LLM] *
# └─ [SPAN] a2a_call_post_reviewer (AutoGen)
# └─ [SPAN] post_reviewer_autogen_execution
# └─ [AGENT] invoke_agent x_post_verifier
# ├─ [LLM] *
# ├─ [Tool] execute_tool character_counter_tool
# ├─ [Tool] execute_tool emoji_detector_tool
# ├─ [Tool] execute_tool post_structure_validator
# ├─ [Tool] execute_tool shakespearean_style_detector
# ├─ [Tool] execute_tool tone_analyzer
# └─ [LLM] *
#
# Notes:
# - Under LangGraph's [Chain] tools there may be batched calls (2 tools can be grouped as 1+1 or 2)
# - AutoGen create_agent nodes are excluded from the reference trajectory
# - AutoGen [SPAN] mcp client/operation and its children are excluded from the reference trajectory
# - Between AutoGen tools there may be 0 or 1 LLM; the evaluator will generate all valid patterns
# ============================================================================
# Default reference trajectory (compatibility): batched LangGraph tools, AutoGen compact mode
reference_trajectory:
# ===== Top-level SPAN =====
- "SPAN: shakespeare_x_post_orchestrator"
# ===== Second-level SPAN (includes retry logic) =====
- "SPAN: crew_execution"
# ===== Stage 1: Topic Analysis (LangGraph; 2 tools) =====
- "SPAN: analyze_topic"
- "SPAN: a2a_call_topic_analyzer"
- "SPAN: topic_analyzer_server_execution"
- "Chain: LangGraph"
- "AGENT: agent"
- "LLM: *"
- "Chain: _should_continue"
- "Chain: tools"
- "Tool: keyword_extractor"
- "Tool: topic_complexity_analyzer"
- "AGENT: agent"
- "LLM: *"
- "Chain: _should_continue"
- "Chain: format_output"
# ===== Stage 2: Content Generation (CrewAI) =====
- "SPAN: content_generation_loop"
- "SPAN: a2a_call_content_generator"
- "SPAN: content_generator_server_execution"
- "Chain: Crew***.kickoff"
- "AGENT: Shakespearean Bard"
- "LLM: *"
- "Tool: keyword_extractor"
- "LLM: *"
- "Tool: character_counter_tool"
- "LLM: *"
- "Tool: emoji_detector_tool"
- "LLM: *"
- "Tool: post_structure_validator"
- "LLM: *"
- "Tool: shakespearean_style_detector"
- "LLM: *"
- "Tool: tone_analyzer"
- "LLM: *"
# ===== Stage 3: Post Review (AutoGen) =====
- "SPAN: a2a_call_post_reviewer"
- "SPAN: post_reviewer_autogen_execution"
- "AGENT: invoke_agent x_post_verifier"
- "LLM: *"
- "Tool: execute_tool character_counter_tool"
- "Tool: execute_tool emoji_detector_tool"
- "Tool: execute_tool post_structure_validator"
- "Tool: execute_tool shakespearean_style_detector"
- "Tool: execute_tool tone_analyzer"
- "LLM: *"
# Target tool list (used for the single-tool use metric)
target_tools:
# Topic Analysis stage (LangGraph)
- "Tool: keyword_extractor"
- "Tool: topic_complexity_analyzer"
# Content Generation stage (CrewAI)
- "Tool: character_counter_tool"
- "Tool: emoji_detector_tool"
- "Tool: post_structure_validator"
- "Tool: shakespearean_style_detector"
- "Tool: tone_analyzer"
# Post Review stage (AutoGen; note the execute_tool prefix)
- "Tool: execute_tool character_counter_tool"
- "Tool: execute_tool emoji_detector_tool"
- "Tool: execute_tool post_structure_validator"
- "Tool: execute_tool shakespearean_style_detector"
- "Tool: execute_tool tone_analyzer"
# 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"
# Permutable tool-group configuration (for dynamic tool-order matching)
permutable_tool_groups:
# Shakespearean Bard validation tools can be in any order (CrewAI part)
# Note: keyword_extractor must be the first tool and is not permuted
shakespearean_bard_validation_tools:
- "Tool: character_counter_tool"
- "Tool: emoji_detector_tool"
- "Tool: post_structure_validator"
- "Tool: shakespearean_style_detector"
- "Tool: tone_analyzer"
# ============================================================================
# Usage notes - A2A_mix hybrid architecture
# ============================================================================
#
# 1. Hybrid framework characteristics:
# - Stage 1 (Topic Analysis): LangGraph
# * Chain name: LangGraph (not Crew***.kickoff)
# * Simplified AGENT name: agent (not a full role name)
# * Batched tool execution: [Chain] tools may contain multiple Tool nodes
# * Special chains: _should_continue and format_output
#
# - Stage 2 (Content Generation): CrewAI
# * Standard CrewAI structure: Crew***.kickoff → AGENT → LLM/Tool
# * Full AGENT name: Shakespearean Bard
# * keyword_extractor must be the first tool; subsequent validation tools may be in any order
#
# - Stage 3 (Post Review): AutoGen
# * Ignore create_agent (excluded from evaluation)
# * Focus on invoke_agent x_post_verifier
# * Tool prefix: execute_tool
# * Filter all [SPAN] mcp client/operation and its children
#
# 2. LangGraph dynamic tool mode (grouping + order):
# The evaluator automatically detects the tools grouping pattern and tool order in each sample,
# and dynamically generates matching reference variants.
#
# Grouping patterns (2 options):
# - [1,1]: 2 Chain: tools, each with 1 Tool
# - [2]: 1 Chain: tools, with 2 Tools
#
# Tool order (2! = 2 options):
# - keyword_extractor → topic_complexity_analyzer
# - topic_complexity_analyzer → keyword_extractor
#
# 3. AutoGen dynamic mode (LLM insertion + tool order):
# The evaluator automatically detects the LLM/Tool calling pattern and tool order in AutoGen,
# and dynamically generates matching reference variants.
#
# LLM insertion patterns (2^4 = 16 options):
# - Compact: LLM → Tool1 → Tool2 → Tool3 → Tool4 → Tool5 → LLM
# (consecutive tool calls with no LLM between tools)
# - Interleaved: LLM → Tool1 → LLM → Tool2 → ... → Tool5 → LLM
# (LLM calls interleaved between tools)
# - Mixed: LLM is inserted between some tools but not others (16 combinations)
#
# Tool order (5! = 120 options):
# - The 5 verification tools can be in any order
# - character_counter_tool, emoji_detector_tool, post_structure_validator,
# shakespearean_style_detector, tone_analyzer
#
# Validation rules:
# - Must start and end with an LLM
# - Must contain all 5 target tools in the middle (order can vary)
# - At most one LLM between adjacent tools
#
# 4. Dynamic reference selection strategy (5-way combination):
# For each sample:
# a) Detect the actual LangGraph tools grouping pattern
# b) Enumerate LangGraph tool order (2! = 2)
# c) Detect the actual AutoGen Post Review LLM/Tool pattern
# d) Enumerate AutoGen tool order (5! = 120)
# e) Enumerate CrewAI tool permutations (5! = 120)
# f) Enumerate all combinations
# - LangGraph grouping: 2
# - LangGraph order: 2
# - AutoGen LLM insertion: 16
# - AutoGen order: 120
# - CrewAI permutations: 120
# - Total variants: 2 × 2 × 16 × 120 × 120 = 921,600
# g) For each variant compute exact_match, in_order_match, any_order_match
# h) Select the variant with the highest composite score
# Composite score = exact_match×3 + in_order_match×2 + any_order_match×1
#
# 5. Metric definitions (consistent with A2A):
# - Exact Match: predicted trajectory matches the reference exactly
# - In-order Match: reference is a subsequence of the predicted trajectory
# - Any-order Match: all required steps are present (order ignored)
# - Precision: fraction of predicted steps that are correct
# - Recall: fraction of reference steps covered by prediction
# - Single-tool Use: average usage rate of target tools
# - unique_path_ratio: path diversity
# - path_entropy: path entropy
#
# 6. Run command:
# cd /Users/wzr/TOSEM-2025/RESULTS/RQ-Failure_Breakdown/SocialMediaManager-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 batched tools execution
# - Dynamic handling of AutoGen LLM/Tool insertion patterns (new)
# - AutoGen tool prefix: execute_tool (not CrewAI tool names)
# - Filter AutoGen MCP client/operation SPAN nodes
# - Triple dynamic matching across LangGraph, CrewAI, and AutoGen
#
# 8. Special filtering rules:
# - Filter all [SPAN] mcp client/operation and its children (AutoGen noise)
# - Filter [AGENT] create_agent (AutoGen initialization, not business logic)
# - Keep [AGENT] invoke_agent (actual AutoGen execution)
#
# 9. Tool-order notes:
# - LangGraph Topic Analysis: 2 tools can be in any order (2! = 2)
# - CrewAI Shakespearean Bard: keyword_extractor must be first; the 5 validation tools can be permuted (5! = 120)
# - AutoGen Post Review: the 5 verification tools can be in any order (5! = 120)
#
|