File size: 16,985 Bytes

8c10cf2

# Full reference trajectory config - BookWriter-MCP project
#
# Version: ideal trajectory derived from the code design (applies to all models)
# Principle: include only steps explicitly required by the design, not based on test statistics
#
# Code design analysis:
# 1. Three main stages executed in order (orchestrator.py):
#    - OutlineGenerator: generate book outline (2 agents: Researcher + Outliner)
#    - WriteBookChapterCrew: write multiple chapters concurrently (2 agents per chapter: Researcher + Writer)
#    - ReviewBookCrew: review the full book (1 agent: Editor)
#
# 2. Tool usage analysis:
#    - OutlineGenerator:
#      - Researcher: bocha_websearch_tool, extract_keywords
#      - Outliner: no tools (reasoning only)
#    - WriteBookChapterCrew (per chapter):
#      - Researcher: bocha_websearch_tool, extract_keywords
#      - Writer: count_words, analyze_chapter_quality, validate_markdown_structure
#    - ReviewBookCrew:
#      - Editor: count_book_words, analyze_book_quality, validate_book_markdown, extract_book_keywords
#
# 3. Execution flow:
#    - orchestrator creates top-level SPAN: book_writing_orchestrator
#    - orchestrator creates crew_execution SPAN with retry logic
#    - three child SPANs: generate_outline, write_chapters, review_book
#    - each Crew contains one Chain and multiple Agents
#    - chapter writing is executed in concurrent batches (batch_size=4)
#
# 4. SPAN hierarchy:
#    - book_writing_orchestrator (top-level, orchestrator.py line 574)
#      └─ crew_execution (contains retry logic, orchestrator.py line 597)
#         ├─ generate_outline (Crew 1, orchestrator.py line 138)
#         ├─ write_chapters (Crew 2 concurrent batches, orchestrator.py line 268)
#         └─ review_book (Crew 3, orchestrator.py line 342)

# Project name
project_name: "BookWriter-MCP"

# Trajectory extraction types
extract_types:
  - "SPAN" # Multi-level SPAN hierarchy
  - "Chain" # kickoff chain for each Crew
  - "Agent" # agents within each Crew
  - "LLM" # LLM calls
  - "Tool" # MCP tool calls

# Repeatable pattern config (for handling varying chapter counts)
# Format: {"start": start step index (0-based), "end": end step index, "min": min repeats, "max": max repeats}
# Reference trajectory indices: 0-11 Outline, 12 write_chapters SPAN, 13-27 single-chapter pattern (full), 28+ Review
repeatable_patterns:
  - start: 13 # Chain: Crew***.kickoff (step 1 of the single-chapter pattern)
    end: 27 # LLM: * (last step of the single-chapter pattern, after validate_markdown_structure)
    min: 3 # at least 3 chapters (business requirement)
    max: 5 # at most 5 chapters
    name: "chapter_writing_pattern" # pattern name (for debugging)

# ============================================================================
# Reference trajectory (ground truth) - derived from the code design
# ============================================================================
#
# Diagram (ideal execution path; the single-chapter pattern repeats 3-5 times):
#
# [SPAN] book_writing_orchestrator
# └─ [SPAN] crew_execution
#    ├─ [SPAN] generate_outline
#    │  └─ [Chain] Crew***.kickoff
#    │     ├─ [Agent] Expert Research Agent for Book Outline Planning
#    │     │  ├─ [LLM] *
#    │     │  ├─ [Tool] bocha_websearch_tool
#    │     │  ├─ [LLM] *
#    │     │  ├─ [Tool] extract_keywords
#    │     │  └─ [LLM] *
#    │     └─ [Agent] Strategic Book Structure Architect
#    │        └─ [LLM] *
#    ├─ [SPAN] write_chapters
#    │  └─ [Repeat the following pattern per chapter, 3-5 times]
#    │     └─ [Chain] Crew***.kickoff
#    │        ├─ [Agent] Specialized Chapter Research Agent
#    │        │  ├─ [LLM] *
#    │        │  ├─ [Tool] bocha_websearch_tool
#    │        │  ├─ [LLM] *
#    │        │  ├─ [Tool] extract_keywords
#    │        │  └─ [LLM] *
#    │        └─ [Agent] Professional Chapter Content Writer
#    │           ├─ [LLM] *
#    │           ├─ [Tool] count_words
#    │           ├─ [LLM] *
#    │           ├─ [Tool] analyze_chapter_quality
#    │           ├─ [LLM] *
#    │           ├─ [Tool] validate_markdown_structure
#    │           └─ [LLM] *
#    └─ [SPAN] review_book
#       └─ [Chain] Crew***.kickoff
#          └─ [Agent] Senior Book Editor and Quality Assurance Specialist
#             ├─ [LLM] *
#             ├─ [Tool] count_book_words
#             ├─ [LLM] *
#             ├─ [Tool] analyze_book_quality
#             ├─ [LLM] *
#             ├─ [Tool] validate_book_markdown
#             ├─ [LLM] *
#             ├─ [Tool] extract_book_keywords
#             └─ [LLM] *
#
# Design basis:
# - orchestrator.py: runs 3 main stages sequentially
# - OutlineGenerator has 2 agents (Researcher + Outliner)
# - WriteBookChapterCrew has 2 agents per chapter (Researcher + Writer)
# - ReviewBookCrew has 1 agent (Editor)
# - all agent roles are defined in each crew's config/agents.yaml
# - tools are obtained from the MCP server via MCPServerAdapter
# - chapter count is produced from the outline, usually 3-5 chapters
#
# Ideal trajectory notes:
# - Top-level SPAN: book_writing_orchestrator
# - Second-level SPAN: crew_execution (contains retry logic)
# - 3 child SPANs run in order: generate_outline → write_chapters → review_book
# - Outline stage: Researcher uses 2 tools; Outliner uses reasoning only
# - Chapter stage: per chapter, Researcher uses 2 tools; Writer uses 3 tools
# - Review stage: Editor uses 4 tools for full-book quality checks
# - This trajectory represents an ideal path with no retries and no redundant steps (4 chapters as a baseline)
#
# Notes:
# - "LLM: *" means any LLM model (wildcard)
# - agent names must match the role field in agents.yaml exactly (full name)
# - tool names must match the MCP server tool names exactly
# - actual chapter count may be 3-5, which affects exact_match but not in-order match
# - tool calls may be skipped or repeated depending on agent decisions
# ============================================================================

reference_trajectory:
  # ===== Top-level SPAN =====
  - "SPAN: book_writing_orchestrator"

  # ===== Second-level SPAN (contains retry logic) =====
  - "SPAN: crew_execution"

  # ===== Stage 1: Generate Outline =====
  - "SPAN: generate_outline"
  - "Chain: Crew***.kickoff"
  # Agent 1: Researcher
  - "Agent: Expert Research Agent for Book Outline Planning"
  - "LLM: *" # decide research strategy
  - "Tool: bocha_websearch_tool" # web search topics
  - "LLM: *" # process search results
  - "Tool: extract_keywords" # extract keywords
  - "LLM: *" # summarize research findings
  # Agent 2: Outliner
  - "Agent: Strategic Book Structure Architect"
  - "LLM: *" # generate the book outline

  # ===== Stage 2: Write Chapters (at least 1; typically 3-5 chapters) =====
  - "SPAN: write_chapters"

  # Standard per-chapter pattern (appears at least once; repeats 3-5 times)
  - "Chain: Crew***.kickoff"
  - "Agent: Specialized Chapter Research Agent"
  - "LLM: *"
  - "Tool: bocha_websearch_tool"
  - "LLM: *"
  - "Tool: extract_keywords"
  - "LLM: *"
  - "Agent: Professional Chapter Content Writer"
  - "LLM: *"
  - "Tool: count_words"
  - "LLM: *"
  - "Tool: analyze_chapter_quality"
  - "LLM: *"
  - "Tool: validate_markdown_structure"
  - "LLM: *"

  # ===== Stage 3: Review Book =====
  - "SPAN: review_book"
  - "Chain: Crew***.kickoff"
  - "Agent: Senior Book Editor and Quality Assurance Specialist"
  - "LLM: *" # start review
  - "Tool: count_book_words" # count words
  - "LLM: *" # analyze word count
  - "Tool: analyze_book_quality" # quality analysis
  - "LLM: *" # assess quality
  - "Tool: validate_book_markdown" # format validation
  - "LLM: *" # check formatting
  - "Tool: extract_book_keywords" # extract keywords
  - "LLM: *" # generate review report

# Target tool list (for the single-tool use metric)
# List all tools that are expected to be used
target_tools:
  # Outline generation stage
  - "Tool: bocha_websearch_tool"
  - "Tool: extract_keywords"
  # Chapter writing stage (per chapter)
  - "Tool: count_words"
  - "Tool: analyze_chapter_quality"
  - "Tool: validate_markdown_structure"
  # Book review stage
  - "Tool: count_book_words"
  - "Tool: analyze_book_quality"
  - "Tool: validate_book_markdown"
  - "Tool: extract_book_keywords"

# ============================================================================
# Dynamic tool permutation configuration
# ============================================================================
# Define permutable tool groups for some agents, to generate reference trajectories with different tool orders.
# During evaluation, all permutations are tried and the best-matching reference is chosen.
#
# Design basis:
# - Chapter Writer's 3 validation tools (count_words, analyze_chapter_quality, validate_markdown_structure)
#   depend only on the chapter draft and are independent of each other, so they can be executed in any order.
# - Book Editor's 4 analysis tools (count_book_words, analyze_book_quality,
#   validate_book_markdown, extract_book_keywords) depend only on book files and are fully independent,
#   so they can be executed in any order.
#
# Number of permutations:
# - Chapter Writer: 3! = 6 permutations
# - Book Editor: 4! = 24 permutations
# - Total: 6 permutations per chapter, and 24 permutations for the Book Editor
#
# Evaluation strategy:
# - For each sample, generate all possible permuted reference trajectories
# - Compute a matching score for each permutation (exact_match * 3 + in_order_match * 2 + any_order_match * 1)
# - Choose the highest-scoring permutation as the reference trajectory for that sample
# - This fairly evaluates models that may call tools in different orders
permutable_tool_groups:
  # Chapter Writer validation tools can be in any order
  # These 3 tools are used within "Agent: Professional Chapter Content Writer"
  chapter_writer_validation_tools:
    - "Tool: count_words"
    - "Tool: analyze_chapter_quality"
    - "Tool: validate_markdown_structure"

  # Book Editor analysis tools can be in any order
  # These 4 tools are used within "Agent: Senior Book Editor and Quality Assurance Specialist"
  book_editor_analysis_tools:
    - "Tool: count_book_words"
    - "Tool: analyze_book_quality"
    - "Tool: validate_book_markdown"
    - "Tool: extract_book_keywords"

# 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 (book_writing_orchestrator)
#    - Level 2: SPAN (crew_execution)
#    - Level 3: SPAN (generate_outline, write_chapters, review_book)
#    - Level 4: Chain (Crew***.kickoff, wildcard for UUID)
#    - Level 5: AGENT (exact match of agent role)
#    - Level 6: LLM (wildcard match for any model)
#    - Level 6: Tool (exact match of tool name)
#
# 2. Wildcard handling:
#    - Crew_<UUID>.kickoff -> Crew***.kickoff (handled automatically by evaluate_trajectory.py)
#    - LLM: * -> matches any model name (e.g., gpt-5-chat-latest, deepseek-reasoner)
#    - Agent names must match exactly (including any prefixes/suffixes)
#
# 3. Exact matching requirements:
#    - SPAN names: "book_writing_orchestrator", "crew_execution", "generate_outline",
#      "write_chapters", "review_book"
#    - Chain name: "Crew***.kickoff" (wildcard match)
#    - Agent names (case-sensitive): "Agent: Expert Research Agent for Book Outline Planning",
#      "Agent: Strategic Book Structure Architect", "Agent: Specialized Chapter Research Agent",
#      "Agent: Professional Chapter Content Writer",
#      "Agent: Senior Book Editor and Quality Assurance Specialist"
#      (must match the role field in agents.yaml exactly)
#    - LLM name: "LLM: *" (wildcard match)
#    - Tool names: must match MCP tool names exactly (bocha_websearch_tool, extract_keywords,
#      count_words, analyze_chapter_quality, validate_markdown_structure,
#      count_book_words, analyze_book_quality, validate_book_markdown, extract_book_keywords)
#
# 4. Metric definitions:
#    - Exact Match: requires identical trajectories (including number of LLM calls; the per-chapter pattern is repeatable)
#    - In-order Match: allows extra calls, but required steps must appear in order
#    - Any-order Match: requires all required steps regardless of order
#    - Precision: fraction of correct steps among predicted steps
#    - Recall: fraction of reference steps covered by predicted steps
#    - Single-tool Use: checks usage of all 9 required tools
#    - unique_path_ratio: path diversity (unique full trajectories / number of samples)
#    - path_entropy: path entropy (Shannon entropy over trajectory frequencies, normalized to 0-1)
#    - Note: chapter count (3-5) does not affect any metric because the per-chapter pattern is repeatable
#
# 5. Run command:
#    cd /Users/wzr/TOSEM-2025/RESULTS/RQ-Failure_Breakdown/BookWriter-MCP
#    python3 evaluate_trajectory.py --config reference_trajectory.yaml
#
# 6. Design notes:
#    - This reference represents an ideal execution path (with a per-chapter pattern, matching 3-5 chapters)
#    - The 3 main stages run in order: Outline → Chapters → Review
#    - Stage 1 (Outline): Researcher uses 2 tools + Outliner uses reasoning only = 4 LLM calls + 2 tool calls
#    - Stage 2 (Chapters): per chapter, Researcher uses 2 tools + Writer uses 3 tools = 8 LLM calls + 5 tool calls per chapter
#    - Stage 3 (Review): Editor uses 4 tools = 5 LLM calls + 4 tool calls
#    - Per-chapter pattern: 1 Chain + 2 Agents + 8 LLM calls + 5 tool calls (repeated within write_chapters)
#    - In-order Match and Recall are often more appropriate for evaluating real performance
#    - Note: the reference defines a single-chapter pattern that repeats 3-5 times to avoid fixed-chapter-count mismatch
#
# 7. Differences vs. actual trajectories:
#    - The reference defines a per-chapter pattern; actual runs repeat it 3-5 times (depending on outline-generated chapters)
#    - Actual trajectories may include more LLM calls (thinking/planning/executing/summarizing; especially for reasoning models)
#    - Some models may call tools multiple times (retries or extra checks)
#    - Writer may repeatedly call count_words or analyze_chapter_quality for iterative improvement
#    - Some models (e.g., Gemini) may skip certain tool calls and still produce results
#    - These differences do not necessarily indicate errors; they reflect different execution strategies
#    - The repeatable per-chapter pattern ensures in_order_match works regardless of the number of chapters
#
# 8. MCP variant characteristics:
#    - All tools are provided by an MCP server over SSE
#    - Tools are retrieved via MCPServerAdapter
#    - Includes a multi-level SPAN hierarchy (orchestrator + crew_execution + 3 stage SPANs)
#    - The 3 stages collaborate to complete the book-writing task
#    - Each stage has clear responsibilities and expected tool usage
#    - OutlineGenerator: 2 agents to research topics and produce an outline
#    - WriteBookChapterCrew: 2 agents to write chapters in concurrent batches
#    - ReviewBookCrew: 1 agent using 4 tools to review full-book quality
#    - Tool categories: search (bocha_websearch_tool), analysis (extract_keywords),
#      quality checks (count_words, analyze_*_quality), format validation (validate_*)
#
# 9. Dynamic tool permutation notes:
#    - Background: based on tool dependency analysis (see tool_sequence_analysis.md)
#      * Chapter Writer's 3 validation tools are independent and can run in any order
#      * Book Editor's 4 analysis tools are fully independent and can run in any order
#    - Implementation:
#      * permutable_tool_groups defines permutable tool groups
#      * Chapter Writer: 3! = 6 permutations
#      * Book Editor: 4! = 24 permutations
#      * Total per sample: 6 × 24 = 144 reference trajectory permutations
#    - Selection strategy:
#      * the evaluation script generates all permutations per sample
#      * computes a matching score for each permutation (exact*3 + in_order*2 + any_order*1)
#      * selects the highest-scoring permutation as that sample's reference
#    - Benefits:
#      * fairly evaluates models that use different tool orders
#      * improves the accuracy of exact_match and in_order_match
#      * reflects realistic flexibility of tool ordering (no forced dependencies)
#    - Performance:
#      * permutation generation and scoring are done in-memory and are efficient
#      * evaluation time per sample increases by ~2-3x (acceptable)