Requests

Inspect a translation stage's model attempts, source batch, raw output, errors, and usage.

A request record describes one model attempt within a translation stage. Use it to investigate a provider failure or unexpected output. A run can contain several documents, locales, and stages, and each stage can make several model requests.

In Studio, open the run, expand the affected document and locale, and select its stage. Open a request from the stage's list to inspect the attempt.

Establish which attempt you are reading #

FieldWhat it tells you
Provider and ModelThe model used for this attempt.
KeyWhether the request used a platform key or an organization connection, when recorded.
StatusWhether this model attempt completed or failed.
Operation, Attempt, and BatchThe operation and batch this attempt belongs to, and its position among attempts.
TokensRecorded input and output token usage for the full request.
Took and StartedThis attempt's duration and start time.
CostRecorded provider cost, when available.

An unsuccessful attempt can be followed by a successful fallback. Compare attempts within the same batch before concluding that a provider error caused the whole target to fail. A model that completed can still return a translation that needs correction.

A request can serve several documents or locales. Its usage and cost describe the full batch, even when you opened it from one target. The same request can therefore appear under multiple targets; count its ID once when combining costs.

Compare the input with the raw output #

Input contains the recorded source batch. Raw output contains the model's response before the rest of the pipeline finishes. Either can be unavailable. A missing payload is not an empty source or proof that the request did no work.

Use the source keys to find the passage being investigated. Compare its meaning, placeholders, and structure with the raw output. If the provider failed, read the error details before changing the content or model selection.

This view does not contain the full system prompt or a complete snapshot of the glossary and rules used for generation. Use the saved source revision, submitted context, and engine configuration history to establish what guidance was available. A glossary's current contents cannot by themselves prove which term an earlier attempt received. Context retrieval explains how guidance is selected.

Follow the result through later stages #

Raw model output is intermediate evidence. Rephrase, human review, post-edit, content-length processing, or back-translation can change the text before the target is reconstructed. Read stage progress, then download the completed target to establish what the caller can consume.

For a review investigation, compare the text offered to the reviewer with the submitted decision, then inspect the final target. A later transformation should not be attributed to the reviewer. Inbox reviews explains the decisions, timeouts, and external-review takeover.

Back-translation uses model-generated reverse text to check for consequential drift and may attempt a correction. It adds model work and can itself be wrong. Read its outcome alongside the original source and final target rather than treating a successful stage as proof of equivalent meaning.

Keep measurements separate from execution #

Radar measurements examine localization events after generation. Their answers can arrive after the target has completed and do not establish whether a model request succeeded. Review Lingo.dev translations explains selection and where to find answers.

Once the original attempt is understood, a reproduction can test a proposed change. Record the new job ID, source revision, engine settings, and output. It is a new execution with possible usage, and later settings can change its result. Debug a bad translation follows the investigation through a repair.