LLM-Wikis
RAG systems with Obsidian that do not forget: four productive knowledge bases, over 5,000 documents.
$ wiki --status
> Knowledge bases running in production
> Typed knowledge graph
> Supersession instead of overwriting
/The problem
Classic RAG fetches context and immediately forgets it again. I wanted knowledge bases that accumulate across sessions, resolve contradictions and keep their sources.
/The system
I build RAG systems with Obsidian: currently 4 productive knowledge bases with over 5,000 ingested documents and transcripts, among them 2,328 YouTube transcripts on SEO and web dev, plus German tax law with citation verification.
Under the surface sit a typed knowledge graph, a confidence decay and supersession instead of overwriting. The ingest pipelines run as a batch over Claude subagents with their own cost tracking.
- Obsidian
- Claude-Subagents
- Typisierter Knowledge Graph
- Confidence-Decay
- Supersession
- Batch-Pipelines + Kostentracking
/Result
- 4
- productive knowledge bases
- 5,000+
- ingested documents
- 2,328
- YouTube transcripts SEO/web dev
The core sentence describes the result best: I build RAG systems that do not forget.
/No interface, just system
$ wiki ingest --batch
> 5,000+ documents and transcripts ingested
> 2,328 YouTube transcripts on SEO and web dev
> 4 knowledge bases kept in sync
> Batch over Claude subagents, cost tracking active
> Confidence decay applied, supersession instead of overwriting
Nothing is overwritten, everything accumulates.
/Stack
- Obsidian
- Claude-Subagents
- Typisierter Knowledge Graph
- Confidence-Decay
- Supersession
- Batch-Pipelines + Kostentracking
/More projects
A similar process with no system yet? I build things like this for others too.