What traffic capture gives you
- Request tracing for every inference that passes through the proxy
- Cost and latency analytics broken down by provider, task, environment, and model
- Datasets created from filtered production traffic
- A direct path into evals and training using the exact requests your product already sees
Two ways to get started
Automatic instrumentation
Use the CLI to scan your codebase and add instrumentation with
inf install.Manual instrumentation
Point your SDK at Inference.net and add the routing and metadata headers yourself.
How it fits into the lifecycle
Capturing traffic is usually the entry point for platform workflows.- Capture real requests.
- Save representative slices as datasets.
- Run evals against the datasets.
- Fine-tune or distill a better model.
- Deploy the result and keep observing production traffic.
What you will work with in the product
Once traffic is flowing, most teams spend time in these surfaces:- Inferences to inspect individual requests and responses
- Tasks to group related traffic by workflow or feature
- Datasets to save the examples you want to evaluate or train on
- Evals to compare models and judge outputs
- Training Jobs to turn good data and good rubrics into a better model
Best fit
Traffic capture is a strong fit when:- you already use OpenAI, Anthropic, or another OpenAI-compatible provider
- you want analytics and trace inspection without rewriting your app
- you want to build evals or training datasets from real production traffic
- you need cleaner visibility into cost, latency, and failures by environment or task
Next steps
Datasets
Save live traffic as datasets or import historical logs.
Evals
Run repeatable quality checks against real datasets.
Talk to an engineer
Meet with our team if you want help planning your eval, training, or migration workflow.