OrionBelt® Semantic Sidecarfor Agentic AI & Knowledge Graphs

Ontology-based MCP server for Text-to-SQL. Analyzes relational schemas, auto-generates RDF/OWL ontologies with embedded SQL mappings, then layers GraphRAG discovery, OBQC validation, and interactive charting on top, so AI agents get relationship-aware, fan-trap-free SQL through any MCP-compatible client.

• 8 database connectors: PostgreSQL, MySQL, Snowflake, ClickHouse, Dremio, BigQuery, DuckDB/MotherDuck, Databricks SQL
• Auto-generates RDF/OWL ontologies from live schemas with oba: namespace SQL annotations and W3C R2RML mappings
• GraphRAG: graph traversal (up to 12 hops) + ChromaDB vector embeddings for semantic schema discovery
• OBQC: deterministic ontology-based SQL validation catches fan-traps, bad joins & type mismatches; errors block execution before queries reach the database
• SPARQL 1.1 query interface via persistent Oxigraph RDF store
• Interactive Plotly charting (bar, line, scatter, heatmap) with MCP-UI rendering in Claude Desktop
• MCP sampling: suggest_semantic_names asks the host LLM to pre-fill business-friendly renames for cryptic identifiers in one call
• 26 MCP tools, multi-schema support with per-schema isolation, and workspace persistence that restores previous sessions
• Pairs with the Semantic Layer: export discovered models as OBML for guaranteed-correct SQL compilation

Open-source Semantic Sidecar for AI, analytics, and governed data systems. Injects governed business semantics into existing platforms with no architecture change and no dedicated semantic infrastructure. Define dimensions, measures, metrics, business rules, and context in declarative YAML (OBML); OBSL compiles them into optimized, dialect-specific SQL and executes via a unified API. Query in business concepts, not raw schemas.

• Semantic Sidecar: drop-in governed semantics for existing platforms, no architecture change
• Analytics as Code: version-controlled YAML compiled into executable SQL, DQ rules, KPIs, and semantic context
• One model powers AI agents, analytics workflows, DQ checks, regulatory & business KPIs, and reporting
• 8 SQL dialects: PostgreSQL, Snowflake, BigQuery, ClickHouse, Databricks, DuckDB/MotherDuck, Dremio, MySQL
• Custom SQL AST for guaranteed correct, injection-safe generation
• Composite Fact Layer (CFL): correct multi-fact, multi-grain queries across separate fact tables with built-in fan-trap and chasm-trap prevention
• Artefacts Composability Resolution (ACR): resolves which dimensions, measures, and metrics stay composable with the current selection, powering guided query builders and safe agent composition
• Agent-facing API: model health on load, query-plan endpoint, structured warning codes
• Freshness Inheritance: cached queries inherit freshness from their source tables, so one heartbeat invalidates every dependent query
• Unified API: REST, MCP server, PostgreSQL Wire Protocol (psql/JDBC/ODBC/BI), Gradio UI, DB-API 2.0 & Apache Arrow Flight SQL drivers, no BI tool in the middle
• Apache Ossie (Open Semantic Interchange) interoperability: bidirectional with Apache Incubator project

Run OBML query batches against the Semantic Layer and emit reports. YAML-defined runs produce self-contained markdown, HTML, or PDF reports with a YAML run-log sidecar capturing compiled SQL, query plans, and timing. Built for cron, CI, and scheduled audits.

• YAML-defined runs: endpoint, named queries, table/value/list sections
• Self-contained markdown, HTML, or PDF reports with locale-aware number formatting
• Always-on YAML run log: compiled SQL, planner explain, server timing, warnings
• Optional per-query TSV exports for downstream tooling
• Three deployment shapes: single-model, multi-model, or runner-loaded session
• Drive from cron, systemd, GitHub Actions, or Cloud Scheduler (no built-in cron)

Browser-based ontology workbench for OWL ontologies and SKOS vocabularies. Streamlit + rdflib. No Java, no Protégé. Built for the people who actually do the modeling work, not just admire the tool.

• Full CRUD for classes, properties, individuals, restrictions, annotations
• Bulk operations: paste names/CSV, spreadsheet edit, multi-select delete
• Validation & OWL-RL reasoning: orphans, duplicate labels, domain/range, SKOS cycles
• Upper-ontology starters (gist by Semantic Arts) and 5 templates
• Merge-aware imports with diff, conflict detection, prefix reconciliation
• Interactive vis-network graph with click-to-edit; full undo/redo

The interface that ties it all together. One conversation to go from raw database schema to ontology, semantic model, compiled SQL, and interactive charts, across cloud or fully local LLMs.

• 300+ models via OpenRouter; Anthropic & OpenAI direct; local via MLX or Ollama
• Dual MCP server support (Analytics + Semantic Layer) with auto-reconnect & retry
• MCP sampling with tools: servers can delegate LLM calls back through the chat client
• Inline Plotly charts (FastMCP ui:// resources) and Mermaid diagrams
• Auto-detect downloadable content: TTL, JSON, CSV, SQL, SPARQL, YAML, XML

X-ray your MCP server: token tax, tool confusion, and surface bloat, distilled into one graded report. Point it at a live server or an offline tools/list dump and walk away with a 0-100 score that answers what the surface costs, whether it confuses the model, and how much smaller it could be.

• One graded report (0-100, letter grade) across five weighted dimensions: context efficiency, selection robustness, surface redundancy, schema hygiene, description quality
• Per-tool context tax via the Anthropic count_tokens endpoint (leave-one-out), so you see what each tool costs before any work
• Behavioral probe: wrong-tool selection and spurious firing, scored against labeled golden queries
• Consolidation analysis: which tools merge, which should be MCP resources, and whether the fix is consolidation or just-in-time loading
• Runs keyless and offline from a tools/list dump, or live over stdio, HTTP & SSE
• Phase-swapped surfaces: per-journey-phase audit where headline tax is the worst phase and progressive loading is credited, not flagged
• Self-contained, replayable run folders, fingerprinted for drift