Ontology-Driven Retrieval Augmented Generation for Real-World Knowledge Graphs

OntoRAG is an open, modular framework that combines ontologies, knowledge graphs, and LLM-based reasoning to build reliable, explainable, and domain-aware Retrieval Augmented Generation systems. It shifts RAG from unstructured text toward semantic, governed, machine-actionable knowledge.

Traditional RAG excels at text recall but struggles with:

• inconsistent schema usage
• domain drift
• poor explainability
• limited structured reasoning
• weak alignment between business definitions and AI usage

OntoRAG addresses this by treating the ontology as the core API, and letting LLMs operate through it rather than around it.

The result: 🔎 more precise retrieval 🧩 consistent reasoning 📚 traceable knowledge ⚙️ LLM-powered workflow automation based on MCP 🛠️ dynamic tools created directly from the graph

OntoRAG is built around three components:

• OWL/RDFS/Base ontology
• Instance graph (RDF)
• Optional provenance and governance layers
• Fully pluggable backend (rdflib, QLever, etc.)

• Document ingestion & ontology induction
• Instance extraction with provenance
• Schema refinement proposals
• Data alignments & semantic diffs
• Multi-step reasoning using current schema cards

OntoRAG generates MCP tools directly from the ontology:

• entity tools (list_X, get_X_by_id, search_X)
• relation tools (get_Y_for_X)
• command tools (close_case, approve_request, …) defined via ontology descriptors

This enables LLM agents to operate on the business knowledge graph with full governance and semantic awareness.

• Ontology induction from documents

• Instance extraction with provenance tracking

• Schema cards for LLM reasoning

• Lightweight or full backend graph engines

  • in-memory RDF (rdflib)
  • QLever high-performance SPARQL backend

• Automatic MCP tool generation from ontology descriptors

• Governance-ready proposal mechanism (alignments, merges, patches)

• You need RAG but your domain has a structure (or should have one).
• You want LLMs to follow domain rules, not invent them.
• You work with business objects (customers, assets, contracts, rules… not just text).
• You want your AI system to be explainable, traceable, and governable.
• You want to integrate LLM reasoning with actual data workflows (via MCP).

Contributions are welcome. We’re particularly interested in:

• ontology engineering tools
• MCP tool generators
• SPARQL adapters & connectors
• document ingestion pipelines
• semantic UI tooling (graph views, diff views)

If you want to get involved, open an Issue or start a Discussion.

TBD — early stage. Likely a permissive license (Apache 2.0 / MIT). Will be finalized before the first public release.

• GitHub Discussions (soon)
• Slack/Discord (coming after first public modules)

Early development Internal prototypes are working; documentation and modularization are in progress. Public components will appear incrementally as they stabilize.