Unit 06.02: Partial answers and how to bound them

Real questions often contain two questions. When the documents answer one of them, an assistant has three options and only one is honest.

Answer the half you have, name the half you do not

The options are: answer both - which requires inventing half; answer neither - which throws away a correct answer; or answer what is supported and state plainly what is not.

The example below takes a compound question, tracks which parts the context supports, and assembles a bounded response.

QUESTION  "What is the refund window, and how do I request one?"

PART OF THE QUESTION     COVERED BY THE DOCUMENTS?
the refund window        yes
how to request one       no

THE RESPONSE
  Refunds are allowed within 7 days. [c1]
  The documents do not cover: how to request one.

  fully answered: no

Answering only the first half without that second sentence is the quiet
failure: the user asked two things and has no signal that one was dropped.

The response gives the refund window with its citation, then adds a sentence naming what the documents do not cover. The fully answered flag is False, which is the field a caller can act on - routing the unanswered half to a human, or logging it as a content gap.

Answering only the first half without that sentence is the quiet failure. The user asked two things, received a confident response, and has no signal that half their question was dropped.

The mistake this prevents

The mistake is treating partial coverage as a binary. Systems built around "answer or refuse" push every compound question into one of the two extremes, and compound questions are common - users bundle related things in one message because that is how people talk.

Takeaway

Track coverage per part of the question, not per question. Answer what is supported, name what is not, and expose a flag so the caller knows the response was incomplete.

Mark complete