dsh-auto-ontology — auto ontology construction with evidence, for DeepSeek Harness
Feed it documents; get a typed ontology (TBox + ABox) where every entity and relation carries its source quote. No evidence, no assertion; unknown types are flagged, never silently dropped. 喂文档,得带类型的本体(TBox+ABox)——每条实体/关系带原文证据。无证据 不落库;未知类型标注拦截,绝不静默丢弃。
Implements the concept of NVIDIA/auto-ontology (13.3k★) natively for DSH, with the family's evidence discipline. Clean-room, zero upstream code. Not affiliated. Family: the 12-repo quant stack + dsh-WrenAI.
Why (ontology plugins exist — none does LLM auto-construction with evidence)
| ontology viewers/bridges (catalog) | dsh-auto-ontology | |
|---|---|---|
| Builds ontologies | no (view/bridge existing) | two-pass LLM pipeline: TBox schema from corpus, then ABox assertions per doc |
| Trust | — | every entity/relation carries an evidence quote; no-evidence assertions don't land |
| Type safety | — | instances must ∈ TBox; unknown types flagged, never dropped |
| Output | files | typed JSON + zero-dep SVG cluster graph + markdown digest |
Quick start
git clone https://github.com/vicfei/dsh-auto-ontology && cd dsh-auto-ontology
python3 scripts/build_ontology.py --docs <dir-or-glob> --domain "<描述>" --save
Real smoke (2026-11-17): fed 5 of our own family READMEs, asked for the
"DSH quant-plugin ecosystem" ontology — got 6 entity types / 49 entities /
45 relations, including (dsh-ai-hedge-fund) -[inspired_by]-> (ai-hedge-fund)
with evidence Inspired by virattt/ai-hedge-fund, and the persona table parsed
into has_role edges. The extractor even kept type discipline: 0 unknown-type
leaks. ontologies/*.json|svg|md carry the full graph with evidence.
Status & roadmap
v0 two-pass engine smoke-tested. ROADMAP: M1 incremental merge + cross-doc coreference; M2 OWL/JSON-LD export + WeKnora graph-KB bridge.
License
MIT; concept credit to Apache-2.0 upstream — zero upstream code used.
No comments yet. Be the first to write one.