Dashboard

Versions

The Versions page lets you compare cost per call across different prompt versions. Tag your deploys with a version label and see exactly how each prompt change affected cost, token usage, and latency — side by side.

How to tag versions

Pass a promptVersion string in your SDK calls. Any string works — a git SHA is the most precise option, but a label like "v2.1" or "nov-rewrite" is easier to read in the dashboard.

await trackedCall({client, params, apiKey, group: 'summarize', userId: session.userId, promptVersion: process.env.DEPLOY_SHA, // or 'v2.1', 'nov-rewrite', etc.})

Once two or more versions have events, the comparison view appears automatically.

What you'll see

The Versions page has two panels:

• Version history — a table of every version label that has events, showing first seen, last seen, total calls, and average cost per call. Sorted most recent first.
• Version comparison — select any two versions to compare them head to head across avg cost per call, avg input tokens, avg output tokens, and avg latency. A delta column shows the percentage change between versions.

What to look for

The most common use case is catching prompt regressions after a deploy. If avg cost per call jumps between versions the likely causes are:

• More input tokens — you added context, examples, or instructions to the system prompt
• More output tokens — the model is generating longer responses, often because the prompt is less specific about output length
• Model change — you switched to a more expensive model for this group

The comparison table breaks out input and output tokens separately so you can identify which side is driving the change.

Using git SHA as version label

Tagging with the deploy SHA gives you the most granular view — every deploy gets its own version automatically with no manual labeling required:

# Set in your deployment environmentDEPLOY_SHA=$(git rev-parse --short HEAD)

// Use in your SDK callsawait trackedCall({... promptVersion: process.env.DEPLOY_SHA,})

If you deploy frequently the version history table can get long. Use descriptive labels like "search-v3" or "gpt4-migration" when you're making a significant prompt change you want to track explicitly.

Next: Budget enforcement →