The problem
Knowledge is spread across runbooks, incident notes and specs. Finding the right paragraph takes longer than acting on it.
How it works
- 01Ingest
- 02Chunk + index
- 03Retrieve
- 04Answer
- 05Cite
What was hard
- Real documents are messy — headers, tables and old copies all hurt search.
- Keeping what the source says separate from what the model guessed.
The goal
Answers come with the paragraph they came from, so people can check the source.
This is a design I worked out on paper — the problem, the approach and the trade-offs. It is not a shipped product.