Notes from building the thing, not from watching the space.

Almost every request that begins we deployed RAG and it does not answer well ends in the same place: the knowledge was never structured. What an ontology actually is, how it differs from a knowledge graph, and a five-step method from terminology control to a measured graph.

Fukko Compass, a disaster-recovery navigation system built by Team Dango. What the legal data made possible, and why the interface mattered as much as the retrieval.

Fixed-length, overlapping, semantic, recursive and structure-based chunking compared across five axes, plus the chunk size that held up across several domains and the overlap ratio beyond which performance degraded.

The blocker is not model accuracy. It is that the business context behind a journal entry was never written down anywhere a machine can read.

Models change every few months. The structure of your data does not. That asymmetry is where a durable advantage comes from, and it is the one most companies leave on the table.

A proof of concept is designed to succeed. Production is designed to survive. Most AI budgets are lost in the gap between those two sentences.

SaaS priced the average of a market. AI makes the specific case cheap to build. That changes what you should buy and what you should own.