Fabric MCP Agent

# Intent Templates - Multi-Stage Execution Framework ## Overview Intent templates provide **domain-agnostic execution patterns** for the multi-stage agentic workflow. These templates define **how** to execute each stage, while persona modules define **what** domain knowledge to apply. ## Template Architecture ### 🎯 Stage 1: Discovery (`stage1_discovery.md`) **Purpose**: Find candidate records using broad search patterns **Generic Approach**: - Initial exploration and filtering - Cast wide net to capture potential matches - Focus on key identifiers and search terms - Applicable to any domain (products, customers, orders, inventory) **Common Patterns**: - Keyword-based searching with LIKE operators - Multiple search criteria with OR logic - Fuzzy matching for product names/descriptions - Category-based filtering **Example Use Cases**: - Product discovery: Find products similar to competitor items - Customer analysis: Identify customers by criteria - Order analysis: Find orders matching patterns - Inventory lookup: Locate parts by specifications ### 📊 Stage 2: Analysis (`stage2_analysis.md`) **Purpose**: Gather comprehensive details on selected candidates **Generic Approach**: - Deep-dive analysis on filtered results from Stage 1 - Join multiple tables for complete picture - Collect detailed attributes, relationships, and metrics - Cross-reference data for validation **Common Patterns**: - Multi-table joins for comprehensive data - Aggregations and calculations - Detailed attribute collection - Relationship mapping between entities **Example Use Cases**: - Product analysis: Get pricing, components, specifications - Customer analysis: Full profile with order history - Order analysis: Complete order details with line items - Inventory analysis: Stock levels, locations, costs ### 🧠 Stage 3: Evaluation (`stage3_evaluation.md`) **Purpose**: Generate business insights and recommendations **Critical Constraint**: **NO SQL EXECUTION** - Pure analysis only **Generic Approach**: - Analyze data patterns from Stages 1 & 2 - Extract key business insights - Generate actionable recommendations - Assess confidence and data quality **Output Structure**: ```json { "business_answer": "Direct answer to user question", "key_findings": ["finding1", "finding2", "finding3"], "recommended_action": "What user should do next", "supporting_data": { "primary_values": "key metrics", "alternatives": "other options", "confidence": "high|medium|low" }, "data_quality": "assessment of result reliability" } ``` ## Template Integration with Personas ### Runtime Combination ``` Intent Template (Generic "How") + Persona Knowledge (Domain "What") = Context-Aware Execution ``` **Example Integration**: - **Stage 1 Template** provides generic discovery patterns - **Product Planning Persona** provides table names (JPNPROdb_pt_mstr) and product-specific logic - **Combined Context** generates product discovery query with domain expertise ### Execution Flow 1. **Intent Classification** determines `multi_stage` strategy 2. **Persona Selection** chooses domain expert (e.g., `spt_sales_rep`) 3. **Template Loading** loads generic stage templates 4. **Context Combination** merges template + persona for each stage 5. **Stage Execution** runs with combined context ## Template Design Principles ### Domain Agnostic - Templates work with any business domain - No hardcoded table names or specific business logic - Generic patterns that adapt to persona context ### Modular and Reusable - Each stage template can be updated independently - New personas automatically benefit from template improvements - Consistent execution patterns across all domains ### AI-Optimized - Designed for LLM consumption and generation - Clear instructions for each stage objective - Structured output requirements for downstream processing ### Extensible Framework - Easy to add new stage types (e.g., `stage4_validation`) - Support for different execution strategies - Future support for iterative and refinement workflows ## Benefits of Template Architecture ### Separation of Concerns - **Templates**: Execution methodology - **Personas**: Domain expertise - **Router**: Strategy selection and orchestration ### Scalability - Add new personas without touching execution logic - Improve execution patterns without updating domain knowledge - Support new domains instantly with existing templates ### Maintainability - Single point of update for execution improvements - Clear responsibility boundaries - Consistent behavior across all personas ### Performance Optimization - Templates can be optimized for SQL generation patterns - Generic patterns enable better caching opportunities - Standardized stage transitions improve predictability ## Future Enhancements ### Advanced Stage Types - **Validation Stage**: Data quality checks and validation - **Refinement Stage**: Iterative query improvement - **Comparison Stage**: Side-by-side analysis patterns ### Conditional Execution - Skip stages based on data quality or completeness - Parallel stage execution for performance - Dynamic stage sequencing based on results ### Template Versioning - Version control for template evolution - A/B testing for template effectiveness - Rollback capabilities for template changes This intent template system provides a robust, scalable foundation for multi-stage business intelligence workflows that can adapt to any domain while maintaining consistent execution patterns.