Run your first analysis
Analysis uses traffic you already captured to look for practical ways to reduce LLM spend. Model and search comparisons are tested against your production answers, so a cheaper option is useful only when it preserves the quality your application already delivers.
This guide assumes you have sent a trace and named at least one route. Analysis is available on the hosted cloud and BYOC deployments. It is not included in the open-source server.
-
Open Find opportunities
Section titled “Open Find opportunities”Open Opportunities and select Find opportunities.
Under What to analyze, let Metergraph select material workloads automatically or choose one workload yourself. Then select the time range. The panel shows the calls and spend available in that window before you start.
-
Choose which changes to try
Section titled “Choose which changes to try”Turn on the changes that make sense for your application:
- Model swap tests the candidate models you select.
- Search swap compares the search options you select.
- Batching is suggested only when results can wait up to the selected Maximum wait.
Review the choices and select Start. Metergraph runs the analysis in the background, and the Opportunities page shows its progress.
-
Read the result
Section titled “Read the result”When the run finishes, the summary shows projected monthly savings and the traffic that was analyzed. It also shows quality and latency changes when those measurements apply.
Each workload row explains what Metergraph found. Model and search opportunities can include quality and latency measurements. A savings opportunity that does not require a newly generated answer, such as result caching, may be marked Not graded and leave quality or latency blank.
-
Review the evidence
Section titled “Review the evidence”Select View evidence before making a change. For a model or search recommendation, you can review the current and proposed setup, projected cost, measured quality and latency, tested alternatives, and individual examples when the report includes them.
For other opportunity types, the evidence focuses on why the workload is eligible and how the saving was estimated.
-
Decide what to change
Section titled “Decide what to change”The recommendation is evidence for a decision, not an automatic production change. Validate it for your application, make the change in your code, and watch the workload as new traffic arrives.
If you use MetergraphBot and the opportunity is supported, Metergraph can prepare a pull request after you approve it. It never merges the change for you.