Vision calls as spans
The instruction, the model, the structured output, the token count and the latency — captured like any other LLM step, so the pipeline has one trace and not two systems.
By agent type · Image
It read the invoice. It read it wrong.
Agents that extract from documents, inspect photos or generate creative have the same quiet-failure problem as text agents and far fewer tools built for them. A vision call is still an LLM call — Netra captures it as a span, with the prompt, the output, the latency and the cost attached.
One instrumentation model, whatever the input
The steps around the model
Pre-processing, OCR, classification, the retry on a bad parse and the downstream tool call all sit in the same span tree as the decision they fed.
Cost per asset
Attribute spend to the document, the render or the route that caused it. Image work is expensive per call and cheap to lose track of.
Extraction, graded
Score structured output against ground truth: field accuracy, JSON validity, refusals and hallucinated values on documents you already know the answer to.
Custom spans where you need them
The SDK is OpenTelemetry-native, so any stage you own — a classifier, a crop, a human review — is one decorator away from being in the trace.
Alerts on drift
Accuracy on a document class slipping after a model swap fires into Slack, with the sessions that caused it attached.
The failures worth naming
Plausible, not present
A vision model asked for a total will usually return a number. Grading against ground truth is the difference between a field that was read and a field that was guessed.
The layout changed
One supplier redesigns an invoice and accuracy drops for that class alone. Clustering by document type makes it a pattern rather than a run of odd tickets.
Quietly expensive
High-resolution retries are the single easiest way to double a bill without changing a line of code. Per-asset cost makes it visible the same week.
Test it against the documents that broke
Turn the assets that failed into a regression set, grade every model or prompt change against it, and keep the same evaluators running on live traffic. Prompt injection hidden in a document is a red-team case, not a surprise.
Start today
Instrument your image pipeline the same way you instrument text
One SDK, one span model, one set of evaluators. Start free.