6.78M
Corpus chunks
Enterprise RAG · knowledge systems
Knowledge the machine returns accurately.
A RAG system’s value rests on retrieval quality and evaluation. I build my own corpus, embedder, reranker and eval harness — not a third-party black box.
Own
Embedder
Own
Reranker
Qdrant
Self-hosted
Hetzner
RAG node
RAGAS
Evaluation
The stack
Four layers, with measurable quality.
Corpus stack
Corpus V3
13.8K books, 6.78M leaf chunks, 31 shards. Own embedder (Qwen3-Emb) and reranker (BGE-v2-m3), self-hosted Qdrant. No compromise on retrieval quality.
Client portal
Natural-language knowledge portal
A client queries its own documents in plain language, with verbatim, attributed quotes. The market-research use is shown separately.
Entity layer
Entity resolution & graph
huSpaCy + GLiNER extraction, Splink-based resolution, an LLMwiki entity hub: knowledge as a cross-referenced entity network, not just text.
Quality
Evaluation built in
RAGAS metrics, golden queries, faithfulness/answeredness gates. The system is done not when it answers — but when it answers verifiably accurately.