Skip to content

Evaluation method

RAG abstention and no-answer testing

How to test what a RAG system does when the answer is not in your documents, when the only source is out of date and when two sources disagree.

Updated 21 Aug 2026

Definition

Abstention is a defined system behaviour under insufficient evidence; it is not a universal promise that a model will know when it is wrong.

01

Define acceptable outcomes first

For each test, specify whether the system should decline, qualify, request clarification, show the conflict or route to a human. Scoring only for the words "I don't know" will miss most of what matters.

  • Decline or qualify
  • Request clarification
  • Show conflict
  • Route to a human
02

Build five failure sets

Build five sets of questions, each one isolating a single reason the evidence is insufficient, so that when it declines you know which one caused it.

  • Answer absent from the document set
  • Only a plausible distractor exists
  • Relevant source is stale
  • Sources conflict
  • Relevant source is outside the user's permission scope
03

Record retrieval and answer separately

Inspect what was retrieved before judging the response. A decline after retrieval failed and a decline because no source exists are different problems.

  • Inspect retrieval first
  • Then judge the response
  • Failed retrieval versus no source
04

Score behaviour

  • Unsupported fact asserted?
  • Appropriate uncertainty expressed?
  • Evidence state exposed?
  • Safe next action recommended?
05

Repeat configuration-sensitive cases

Model, prompt, retrieval settings and document set updates can change behaviour. Record configuration and rerun critical cases rather than treating one response as stable.

  • Model and prompt changes
  • Retrieval and document set updates
  • Record the configuration
  • Rerun critical cases

What this page does not prove

  1. B1No-answer tests reduce risk but do not eliminate unsupported answers.
  2. B2A refusal can also be wrong when adequate evidence exists.
  3. B3Permission-boundary tests must avoid exposing restricted content in logs or fixtures.
  4. B4Publish results only with method and configuration.

Test the claim on your documents

Pick a real piece of work, agree what a good answer looks like, then go through the results together.