Ontology-Based NL SearchENVITED-X Simulation Assets

Translating natural language into precise, ontology-grounded SPARQL queries — with full transparency about what the AI understood and what it couldn't.

View Presentation →

Architecture

🗣️ Natural Language Input

Users describe what they need in plain language — in any language. No SPARQL or ontology knowledge required.

🧠 Ontology-Grounded AI

OWL + SHACL ontologies are loaded at startup; raw SHACL shapes ground the LLM while graph-driven schema queries feed the compiler.

✅ Post-LLM Validation

A slot validator corrects filter values, applies multi-domain-aware domain correction, and recomputes confidence against the real ontology.

🎯 Deterministic SPARQL

LLM fills structured slots — a deterministic compiler generates verified SPARQL. No hallucinated queries, ever.

🌐 Graph-Driven Multi-Domain Search

Shared properties such as `roadTypes` can span HD maps, OSI traces, scenarios, environment models, and surface models without hardcoded domain tables.

🧭 Traceability + Lineage (WP3)

Every result row carries a per-row predicate-chain breadcrumb back to source; the "Explore lineage" panel walks outgoing references multi-hop across asset classes.

⚡ Progressive Streaming

Results stream via SSE as each pipeline phase completes. Users see the AI interpretation immediately while execution runs.

🔌 Multi-Provider

5 providers via Vercel AI SDK (OpenAI, Anthropic, claude-cli, vibe-cli/Mistral, Ollama) plus GitHub Copilot SDK. Same validation pipeline, different LLM backends.