Grounded Answers From Documents
The questions that stump your best people live in documents nobody reads. Build one grounded question-answering capability over a governed corpus - and know exactly what its answers do and do not prove.
6 modules · 22 lessons · ~6.7 hours · Analysts, domain experts, records owners, and the engineers who support them
Chapter 6 in the Meridian sequence · 9 live so far · builds on Building Trustworthy Data Products and Zero Trust Implementation · continues in AI Assurance: System Risk and Release Decisions
Meridian Utilities can prove its numbers now. What it cannot answer is the question the planning supervisor asked last month: "we replaced that regulator bank in 2019 — what did the crew find inside?" The answer exists. It is on page 40 of a maintenance manual, in a field note from a truck, in a filing from the interconnect contract — documents nobody has read in years. This training builds one grounded question-answering capability over a governed corpus of exactly those documents, and it is honest about the hard part: not getting an answer, but knowing what the answer proves. You will write the evaluation questions before anything retrieves, govern what the system is allowed to read, build the retrieval with an AI CLI, check answers against their cited sources by hand, and retire the whole thing the day it stops earning its keep. Every lesson uses AI CLI tools to do the real work — and the training's own thesis is deliberately modest: a grounded answer converts "is this true?" into "does this passage say this?" — a check you can actually perform. No prior Meridian knowledge needed.
The Curriculum
Decide the Answering Method
A bounded class of questions, not a chatbot
The questions that live in documents, the four ways to answer them that need no build at all, and the question set frozen before anything retrieves
Govern What It Reads
The corpus is a governed dependency
What goes in and what never does, the extraction step where real corpora die, permissions designed before retrieval exists, and a corpus that knows its own version
Retrieve the Evidence
Find it, or say so
The retrieval contract, one quarantined build lesson, permissions enforced where the evidence is fetched, and the eval that asks whether the evidence was found at all
Prove the Answer
Support is not truth
The hand check that citations demand, abstention as a feature, the routing table for which half failed, and the honest limits of the whole approach
Authorize the Answer
Authorize the answer, not just the source
Answers that aggregate past any passage, the question-and-answer record as a corpus of its own, and what it costs to ask
Re-test Every Changed Dependency
Any change re-runs the evals
The answer record that names its versions, the changes that silently invalidate old evaluations, and the ownership that decides when to shut it down
New to AI CLI tools?
This training assumes you can drive an AI CLI (Claude Code, Codex CLI, Antigravity CLI, or Copilot CLI). If that's new, these modules from our AI-Powered Development training are the fastest preparation — most students need only the first one: