The problem

The client runs courses on an LMS that has been in production for years. After each live session, the discussion board fills with questions the lesson already answered: where the reading said it, what the example was asking. Instructors were replying one by one. The replies were short. The volume was the cost.

They did not want a tutor that could roam the open web or invent a worked solution the lesson had withheld on purpose. The answer had to be a passage from the materials the student was already assigned, or a clear decline.

Why the lesson is the boundary

An answer from outside the lesson creates two problems. It can contradict the instructor, and it can give away a graded question the materials deliberately did not solve. The product boundary is the assigned materials for that unit. Anything else is a decline that points the student back to the instructor.

Architecture

Retrieval is the unit, not the catalog

Each course unit has a materials list in the LMS. The assistant searches that list only. A better-matching paragraph from a later unit, or from another course, is out of scope even if it is a better answer.

The passage is the answer

The released reply quotes the passage and names the material. A paraphrase with no passage is rejected by the same check. Students can open the material from the reply.

Graded prompts are excluded

The client marked which prompts are graded and must not be solved in the board. Those threads get a decline even if a passage would hint the solution. The instructor still sees the question.

It sits beside the board

We did not replace the discussion product. The assistant drafts a reply the instructor can post, edit, or discard. In the sections that opted in, a passage-backed reply can post as the course assistant, still with the instructor able to remove it.

A design decision worth calling out
We kept instructor posting as the default and made automatic posting an opt-in per section. A program lead wanted it on for every course in week two. The instructors who had not reviewed the decline rules were the ones who needed the default. Opt-in kept the pilot honest.

Challenges and tradeoffs

Results

Across the opted-in sections, instructor time on questions the client classified as repeats fell 46%, using the time sample the program already collected. Every reply the assistant released included a passage from that unit. Reviewers did not find a released answer that solved a graded prompt or cited material from outside the unit.

Students with a question the lesson does not answer still reach the instructor. The board is quieter where the lesson already spoke.

What we'd do differently

We would have built the graded-prompt paraphrase check before the first opted-in section, not after the first near-miss an instructor caught.