| # Complete Reference Trajectory Configuration - SQLAssistant-MCP Project | |
| # | |
| # Version: Ideal trajectory based on code design (applicable to all models) | |
| # Principle: Only include steps explicitly required by code design, not based on test statistics | |
| # | |
| # Code Design Analysis: | |
| # 1. Execution Flow (orchestrator.py): | |
| # - Orchestrator manages entire workflow with two-level retry logic: | |
| # (a) Orchestrator-level retry: For exception errors (max 3 attempts, MAX_CREW_RETRIES=2) | |
| # (b) Business logic retry: For compliance check failures (max 3 attempts, max_business_retries=2) | |
| # - Ideal flow (no retry): | |
| # Step 1-2: Generate and review SQL (SQLGenerationCrew) | |
| # Step 3: Compliance check (ComplianceCheckerCrew) | |
| # Step 4: Execute SQL query (only if compliance passes) | |
| # Step 5: Interpret results (ResultInterpreterCrew, only if compliance passes) | |
| # | |
| # 2. Three Crew Agent Structure: | |
| # - SQLGenerationCrew (sql_generation/sql_generation_crew.py): | |
| # - Agent 1: Expert SQL Query Generator | |
| # Tools: get_database_schema, get_table_sample, get_column_stats, list_tables | |
| # - Agent 2: Senior SQL Code Reviewer | |
| # Tools: validate_sql_syntax, get_database_schema, check_table_exists | |
| # - ComplianceCheckerCrew (compliance_checker/compliance_checker_crew.py): | |
| # - Agent: Data Security and Compliance Auditor | |
| # Tools: get_database_schema, check_table_exists | |
| # - ResultInterpreterCrew (result_interpreter/result_interpreter_crew.py): | |
| # - Agent: Business Intelligence Analyst | |
| # Tools: run_sql_query, count_rows | |
| # | |
| # 3. SPAN Hierarchy: | |
| # - sql_assistant_workflow (top level, orchestrator.py line 226) | |
| # - crew_execution (contains retry logic, orchestrator.py line 243) | |
| # - generate_sql (SQL generation, orchestrator.py line 266) | |
| # - check_compliance (compliance check, orchestrator.py line 316) | |
| # - interpret_results (result interpretation, orchestrator.py line 380, only if compliance passes) | |
| # | |
| # 4. MCP Tools (tools/mcp_server.py): | |
| # - Database exploration: get_database_schema, get_table_sample, get_column_stats, list_tables | |
| # - SQL validation: validate_sql_syntax, check_table_exists | |
| # - Query execution: run_sql_query, count_rows | |
| # Project name | |
| project_name: "SQLAssistant-MCP" | |
| # Trajectory extraction type configuration | |
| extract_types: | |
| - "SPAN" # Multi-level SPAN structure | |
| - "Chain" # Each Crew's kickoff chain | |
| - "Agent" # Each Crew's Agent | |
| - "LLM" # LLM calls | |
| - "Tool" # MCP tool calls | |
| # Repeatable pattern configuration | |
| # SQLAssistant-MCP has no repeatable patterns (unlike book writing with multiple chapters) | |
| repeatable_patterns: [] | |
| # ============================================================================ | |
| # Reference Trajectory (Ground Truth) - Based on Code Design | |
| # ============================================================================ | |
| # | |
| # Graphical Structure (ideal execution path, no retry, no error, each Agent uses | |
| # at least the key tools required in config/tasks; other tools are optional): | |
| # | |
| # [SPAN] sql_assistant_workflow | |
| # +-- [SPAN] crew_execution | |
| # +-- [SPAN] generate_sql | |
| # | +-- [Chain] Crew***.kickoff | |
| # | +-- [Agent] Expert SQL Query Generator | |
| # | | +-- [LLM] * | |
| # | | +-- [Tool] get_database_schema | |
| # | | +-- [LLM] * | |
| # | +-- [Agent] Senior SQL Code Reviewer | |
| # | +-- [LLM] * | |
| # | +-- [Tool] validate_sql_syntax | |
| # | +-- [LLM] * | |
| # | +-- [Tool] get_database_schema | |
| # | +-- [LLM] * | |
| # +-- [SPAN] check_compliance | |
| # | +-- [Chain] Crew***.kickoff | |
| # | +-- [Agent] Data Security and Compliance Auditor | |
| # | +-- [LLM] * | |
| # | +-- [Tool] get_database_schema | |
| # | +-- [LLM] * | |
| # +-- [SPAN] interpret_results | |
| # +-- [Chain] Crew***.kickoff | |
| # +-- [Agent] Business Intelligence Analyst | |
| # +-- [LLM] * | |
| # +-- [Tool] run_sql_query (execute query) | |
| # +-- [LLM] * | |
| # | |
| # Design Basis: | |
| # - orchestrator.py: Sequential execution of 5 steps (generate, review, compliance, execute, interpret) | |
| # - SQLGenerationCrew contains 2 Agents, sequential execution (Generator -> Reviewer) | |
| # - ComplianceCheckerCrew contains 1 Agent | |
| # - ResultInterpreterCrew contains 1 Agent, uses run_sql_query to execute query | |
| # - All Agent roles defined in each crew's config/agents.yaml | |
| # - Tools obtained from MCP server via MCPServerAdapter | |
| # | |
| # Ideal Trajectory Notes: | |
| # - Top-level SPAN: sql_assistant_workflow | |
| # - Second-level SPAN: crew_execution (contains two-level retry logic) | |
| # - 3 sub-SPANs execute sequentially: generate_sql -> check_compliance -> interpret_results | |
| # - generate_sql stage: Generator explores schema, Reviewer validates syntax | |
| # - check_compliance stage: Auditor checks security and compliance, must pass (PASS verdict) | |
| # - interpret_results stage: BI Analyst uses run_sql_query to execute and interpret results | |
| # - This trajectory represents ideal execution path with no retry, no error | |
| # | |
| # Notes: | |
| # - "LLM: *" means any LLM model (wildcard match) | |
| # - AGENT names exactly match role field in agents.yaml | |
| # - Tool names exactly match MCP server provided tool names | |
| # - Tool calls may vary based on Agent decisions (e.g., Generator may use get_database_schema or list_tables) | |
| # - Reviewer may skip tool calls (if SQL is already correct) | |
| # - Auditor may skip tool calls (if no additional verification needed) | |
| # ============================================================================ | |
| reference_trajectory: | |
| # ===== Top-level SPAN ===== | |
| - "SPAN: sql_assistant_workflow" | |
| # ===== Second-level SPAN (contains two-level retry logic) ===== | |
| - "SPAN: crew_execution" | |
| # ===== Stage 1: Generate SQL (SQLGenerationCrew) ===== | |
| - "SPAN: generate_sql" | |
| - "Chain: Crew***.kickoff" | |
| # Agent 1: SQL Generator (uses key tools required in config/tasks) | |
| - "Agent: Expert SQL Query Generator" | |
| - "LLM: *" # Decide overall generation strategy | |
| - "Tool: get_database_schema" # Per tasks.yaml, first view complete schema | |
| - "LLM: *" # Generate candidate SQL query based on schema | |
| # Other tools (list_tables / get_table_sample / get_column_stats) are optional exploration tools | |
| # Agent 2: SQL Reviewer (uses key tools required in config/tasks) | |
| - "Agent: Senior SQL Code Reviewer" | |
| - "LLM: *" # Read and understand generated SQL | |
| - "Tool: validate_sql_syntax" # Per tasks.yaml and backstory, must check syntax | |
| - "LLM: *" # Analyze syntax check results | |
| - "Tool: get_database_schema" # Per tasks.yaml, verify tables and columns exist | |
| - "LLM: *" # Correct SQL based on schema check results | |
| # ===== Stage 2: Check Compliance (ComplianceCheckerCrew) ===== | |
| - "SPAN: check_compliance" | |
| - "Chain: Crew***.kickoff" | |
| - "Agent: Data Security and Compliance Auditor" | |
| - "LLM: *" # Initial SQL reading and risk identification | |
| - "Tool: get_database_schema" # Required: view schema to assess table/field sensitivity | |
| - "LLM: *" # Combine SQL text and schema to give PASS/FAIL verdict | |
| # ===== Stage 3: Interpret Results (ResultInterpreterCrew) ===== | |
| # Note: This stage only executes if compliance passes | |
| - "SPAN: interpret_results" | |
| - "Chain: Crew***.kickoff" | |
| - "Agent: Business Intelligence Analyst" | |
| - "LLM: *" # Understand business question and reviewed SQL | |
| - "Tool: run_sql_query" # Per tasks.yaml and backstory, must use run_sql_query to get real data | |
| - "LLM: *" # Generate business interpretation based on real results (count_rows etc. are optional) | |
| # Target tools list (for single-tool use metric) | |
| # Lists all tools filtered for each Agent in configuration | |
| target_tools: | |
| # SQL Generation stage - Generator | |
| - "Tool: get_database_schema" | |
| - "Tool: get_table_sample" | |
| - "Tool: get_column_stats" | |
| - "Tool: list_tables" | |
| # SQL Generation stage - Reviewer | |
| - "Tool: validate_sql_syntax" | |
| - "Tool: check_table_exists" | |
| # Compliance stage | |
| # (shares get_database_schema and check_table_exists with Reviewer, not repeated) | |
| # Result Interpretation stage | |
| - "Tool: run_sql_query" | |
| - "Tool: count_rows" | |
| # 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 Instructions | |
| # ============================================================================ | |
| # | |
| # 1. Trajectory Hierarchy: | |
| # - Level 1: SPAN (sql_assistant_workflow) | |
| # - Level 2: SPAN (crew_execution) | |
| # - Level 3: SPAN (generate_sql, check_compliance, interpret_results) | |
| # - Level 4: Chain (Crew***.kickoff, wildcard handles UUID) | |
| # - Level 5: AGENT (exact match agent role) | |
| # - Level 6: LLM (wildcard matches any model) | |
| # - Level 6: Tool (exact match tool name) | |
| # | |
| # 2. Wildcard Handling: | |
| # - Crew_<UUID>.kickoff -> Crew***.kickoff (auto-handled by evaluate_trajectory.py) | |
| # - LLM: * -> matches any model name (e.g., gpt-5-chat-latest, deepseek-reasoner) | |
| # - AGENT names must match completely (including full role name) | |
| # | |
| # 3. Exact Match Requirements: | |
| # - SPAN names: "sql_assistant_workflow", "crew_execution", "generate_sql", | |
| # "check_compliance", "interpret_results" | |
| # - Chain names: "Crew***.kickoff" (wildcard match) | |
| # - Agent names: "Agent: Expert SQL Query Generator", | |
| # "Agent: Senior SQL Code Reviewer", | |
| # "Agent: Data Security and Compliance Auditor", | |
| # "Agent: Business Intelligence Analyst" | |
| # (exact match role field in each agents.yaml) | |
| # - LLM names: "LLM: *" (wildcard matches any model) | |
| # - Tool names: exact match MCP tool names (get_database_schema, list_tables, | |
| # get_table_sample, get_column_stats, validate_sql_syntax, | |
| # check_table_exists, run_sql_query, count_rows) | |
| # | |
| # 4. Evaluation Metric Meanings: | |
| # - Exact Match: Requires identical trajectory (including LLM and tool call counts) | |
| # - In-order Match: Allows extra calls, but core steps must appear in order | |
| # - Any-order Match: Only requires all necessary steps present (ignores order) | |
| # - Precision: Proportion of correct steps in predicted trajectory | |
| # - Recall: Proportion of reference steps covered | |
| # - Single-tool Use: Detects core tool usage | |
| # - unique_path_ratio: Path diversity (unique complete trajectories / sample count) | |
| # - path_entropy: Path entropy (Shannon entropy based on trajectory frequency, normalized to 0-1) | |
| # | |
| # 5. Run Command: | |
| # cd SQLAssistant-MCP | |
| # python3 evaluate_trajectory-Filter_Tools.py --config reference_trajectory.yaml | |
| # | |
| # 6. Design Notes: | |
| # - This reference trajectory represents ideal execution path (no retry, no error) | |
| # - 3 main stages execute sequentially: SQL Generation -> Compliance Check -> Result Interpretation | |
| # - Stage 1 (SQL Generation): Generator uses tools to explore schema + Reviewer validates syntax | |
| # - Stage 2 (Compliance Check): Auditor checks security, must pass | |
| # - Stage 3 (Result Interpretation): BI Analyst uses run_sql_query to execute and interpret | |
| # - In-order Match and Recall metrics are more suitable for evaluating actual performance | |
| # - Reference trajectory only defines core required steps, allows extra tool calls and LLM reasoning | |
| # | |
| # 7. Differences from Actual Trajectories: | |
| # - Actual trajectories may contain business logic retries (compliance check failure triggers SQL regeneration) | |
| # - Actual trajectories may contain orchestrator-level retries (exception errors trigger workflow retry) | |
| # - Generator may use multiple tools to explore database (get_database_schema, list_tables, etc.) | |
| # - Reviewer may call validate_sql_syntax multiple times for syntax validation | |
| # - Auditor may use get_database_schema to verify table and column existence | |
| # - BI Analyst may use count_rows to verify result count | |
| # - Reasoning models (e.g., DeepSeek-R1) may include many extra LLM thinking calls | |
| # - Some models may skip certain tool calls and generate results directly | |
| # - These differences don't represent errors, but different model execution strategies | |
| # | |
| # 8. MCP Version Features: | |
| # - All tools provided via MCP server using SSE protocol | |
| # - Tools obtained from server via MCPServerAdapter | |
| # - Contains multi-level SPAN structure (orchestrator + crew_execution + 3 stage SPANs) | |
| # - 3 specialized Crews collaborate to complete SQL query task | |
| # - Each Crew has clear responsibilities and available tool sets | |
| # - SQLGenerationCrew: 2 Agents, generate and review SQL | |
| # - ComplianceCheckerCrew: 1 Agent, check security and compliance | |
| # - ResultInterpreterCrew: 1 Agent, execute query and generate business insights | |
| # - Tool categories: exploration tools (get_database_schema, list_tables, get_table_sample, etc.), | |
| # validation tools (validate_sql_syntax, check_table_exists), | |
| # execution tools (run_sql_query, count_rows) | |
| # | |
| # 9. Business Logic Notes: | |
| # - Compliance check is critical: only if compliance passes will query execute | |
| # - If compliance fails, SQL is regenerated (max 3 attempts) | |
| # - If all attempts fail, workflow ends without executing query | |
| # - interpret_results stage only appears after compliance passes | |
| # - Therefore, some failed trajectories may lack interpret_results stage | |