Semantic Layer Roadmap

[cjimti]

Semantic Layer Roadmap

Tracking issue for semantic layer and MCP Apps enhancements.

Audience

Primary: Management and Executives

• Furthest from raw data, highest need for self-service
• Don't know (and shouldn't need to know) SQL, tables, columns, lineage
• Need accurate, insightful, visually compelling answers to business questions

Secondary: Data Analysts and Technicians

• Benefit from the same tools with ability to drill into technical details
• Technical metadata builds credibility ("this isn't magic")

Design Principles

1. Insights over mechanics - Lead with the answer/visualization, technical details in the background
2. Platform owns presentation - Toolkits (mcp-trino, mcp-datahub, mcp-s3) are building blocks; mcp-data-platform crafts the user experience
3. MCP Apps live here - Apps need cross-injection, semantic enrichment, unified design language
4. Enrichment over tools - Prefer middleware that enhances existing tool responses over adding new tools
5. Credibility through transparency - Subtle technical details (query time, row count) build trust without overwhelming

Architecture

graph TB
    subgraph Platform["mcp-data-platform"]
        subgraph Apps["MCP Apps Layer"]
            QR[Query Results]
            LV[Lineage Viz]
            CB[Catalog Browser]
        end
        
        subgraph Enrichment["Semantic Enrichment"]
            CI[Cross-Injection]
            BC[Business Context]
            DQ[Data Quality]
        end
        
        subgraph Toolkits["Composed Toolkits"]
            Trino[mcp-trino]
            DataHub[mcp-datahub]
            S3[mcp-s3]
        end
    end
    
    User([Executive / Analyst]) --> Apps
    Apps --> Enrichment
    Enrichment --> Toolkits
    Toolkits --> Data[(Data Sources)]
    
    style Apps fill:#3b82f6,color:#fff
    style Enrichment fill:#8b5cf6,color:#fff
    style Toolkits fill:#6b7280,color:#fff

Loading

Toolkits expose data. Platform presents insights.

Current State

• ✅ Semantic enrichment middleware (adds DataHub context to Trino responses)
• ✅ Cross-injection pattern (Trino ↔ DataHub bidirectional context)
• ✅ Query Results MCP App
  • Auto-chart for numeric data (insight first)
  • Interactive table with sort/filter/search
  • Smart titles ("Revenue by Product" not "column1 by column2")
  • Subtle footer stats (credibility without noise)
  • CSV export

Potential Enhancements (prioritize as needed)

MCP Apps

• Catalog Browser - Searchable data catalog for discovery ("what data do we have about customers?")
• Lineage Visualization - Interactive graph showing data flow
• Quality Dashboard - Data freshness, quality scores at a glance
• Schema Explorer - Visual schema with business descriptions

Middleware Enrichment

• Include data quality scores in responses
• Add deprecation warnings prominently
• Surface freshness info (last updated)
• Inline glossary term definitions

Query Results App Improvements

• Multiple chart support (compare metrics)
• Sparklines for trend data
• Conditional formatting (highlight outliers)
• Saved views / bookmarks

Dependencies

• QueryStats.Duration JSON field misleading: named duration_ms but contains nanoseconds mcp-trino#22 - Fix duration_ms serialization
• Enhancement: Add query_id to QueryStats for debugging/tracking mcp-trino#23 - Add query_id to response
• Enhancement: Include Trino warnings in query response mcp-trino#24 - Include Trino warnings

Non-Goals

• Adding many new tools (tool sprawl)
• Exposing technical metadata to end users
• Duplicating DataHub/Trino functionality