How should an application handle refusals, incomplete output, and truncation?

Separates provider refusals, safety blocks, token truncation, transport failure, and valid domain abstention.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

I inspect the provider’s response status before parsing the success schema. A refusal or safety block, token-limit truncation, timeout, cancelled stream, and valid insufficient-evidence result are different outcomes. I map each one to an explicit application state and a useful user response.

Full answer below

Practice focus

Whether distinct model and transport outcomes receive distinct application behavior.

Answer first

Answer path

Core answer first, then checkpoints toward a stronger answer.

Start
Core answer
Start here. This is the direct spoken answer to practice first.
Overview
A missing success payload is not one generic parsing error. The reason determines whether to retry, ask for different input, fall back, or stop.
I inspect the provider’s response status before parsing the success schema. A refusal or safety block, token-limit truncation, timeout, cancelled stream, and valid insufficient-evidence result are different outcomes. I map each one to an explicit application state and a useful user response.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
Truncation may justify reducing context, increasing an allowed output budget, or continuing only if the protocol supports safe continuation. A refusal normally should not be blindly retried with the same request. A network timeout may be retryable, but only within a deadline and with idempotent downstream behavior. I keep the original failure classification in telemetry.

Review signals

Use these to catch weak phrasing, shallow reasoning, and risky claims.

Common mistakesCandidateCommon mistakesMistakes that weaken the answer or signal shallow understanding.Open details

Details

Parsing every non-success response as malformed JSON.
Retrying a refusal repeatedly with identical input.
Showing a partial structured object as a complete result.
Hiding provider stop reasons behind one generic error.

How should an application handle refusals, incomplete output, and truncation?

Practice focus

Answer path

Overview

Review signals

Follow-up prompts

Continue practice

Go deeper

Tech stacks

Skill areas

Human editorial direction

Was this drill useful?

Found this answer useful?