Knowledge systems

Oghma Infinium

An experimental, local-first RAG workbench for building, inspecting, and measuring retrieval machinery instead of treating it as a black-box stack.

ExperimentalLocal-first RAGEvaluation work

The project

Retrieval is an engineering problem, not a checkbox.

Oghma Infinium explores the parts underneath a useful answer: ingest and chunking decisions, parent-child relationships, embeddings and vector search, lexical BM25 search, and hybrid retrieval where each signal can compensate for the other.

It makes the surrounding work visible too. Query expansion, cross-encoder reranking, grounding checks, gold-set evaluation, latency, index health, and observability are treated as things to understand and measure, not jargon to paste onto a feature page.

  • Ingest and chunking experiments
  • Parent-child retrieval strategies
  • Embedding, vector, and lexical retrieval
  • Hybrid search and query expansion
  • Cross-encoder reranking
  • Grounding, evaluation, and index observability
Current state

Experimental. The point is to learn which retrieval decisions hold up under measurement before calling the stack solved.