A hybrid GraphRAG system built from personal Freeplane mindmaps. Combines deterministic SPARQL querying over an RDF knowledge graph with semantic vector search via LanceDB to enable grounded, natural-language Q&A over a personal knowledge base.

Status: Weeks 1–7 of 12 complete · Active development

Freeplane .mm files (10 maps)
        │
        ▼
parse_mm_to_rdf.py          ← .mm XML → RDF triples (rdflib)
        │
        ▼
outputs/*.ttl               ← 142,796 triples across 10 maps
        │
   ┌────┴────┐
   ▼         ▼
validate_rdf.py        embed_to_lancedb.py
(SPARQL queries)       (fastembed → LanceDB)
   │                         │
   ▼                         ▼
pkg_store/             pkg_lancedb/
(RDF triple store)     (31,983 vectors, 384-dim)
        │
        ▼
retrieve.py                 ← Hybrid retrieval (SPARQL + LanceDB)
        │
        ▼
   [Week 8] Claude API → grounded Q&A

Total: 31,983 embedded concepts · 142,796 RDF triples

pitchstone.mm and neogov.mm are permanently excluded (employer-proprietary data).

Namespace: https://pkg.chunnodu.com/ontology#

Built on standard vocabularies — skos: for concept hierarchy, schema: for typed resources, dc: for metadata — with a minimal custom pkg: namespace for project-specific types and properties.

Key custom types: pkg:Task, pkg:PersonalNote, pkg:Resource, pkg:LogEntry, pkg:Goal, pkg:Project

Key custom properties: pkg:hasSubTopic, pkg:sourceMap, pkg:status, pkg:dateLogged

See pkg_ontology.ttl for the full schema.

• Python 3.10+ · rdflib · fastembed · lancedb · pyarrow
• Embeddings: BAAI/bge-small-en-v1.5 (384-dim, ONNX via fastembed)
• Vector DB: LanceDB (embedded, no server)
• RDF: Turtle serialisation, SPARQL via rdflib
• Source format: Freeplane .mm (XML)