The waste detector

Every captured call is run through a waste detector. It produces concrete findings, each carrying how much it could reclaim, and rolls them up into a single headline: how much of your spend is reclaimable.

What counts as reclaimable

llmprof is deliberate about not overstating savings. It only counts spend you can actually recover, priced at the call’s cache-aware rate so the figure never exceeds what you paid:

• Duplicated content: the same block (a RAG chunk, an instruction) appearing more than once in the context. The extra copies are removable tokens.
• Uncached stable prefix: a system prompt + tool schemas that repeat on every call. Prompt caching can cut about 90% off them after the first call (a recurring saving, not removed tokens).
• Tools never used across the window: schemas for tools you ship on every call but that are never invoked across all captured calls. This is gated - a tool unused on a single call is not waste (an agent needs its full toolset across a run), so llmprof waits until it has seen enough calls before claiming it.

Advisory findings

Some findings flag a smell without claiming a precise reclaimable amount, because the right fix depends on your app, or because acting on them is not always possible:

• Tools not called on a given request (informational; only tools never used across the whole run are counted as reclaimable).
• Tool schemas are a large share of the context (trim descriptions, lazy-load).
• History and tool results dominate (summarize or truncate older turns).
• A very large system prompt riding every call.

The monthly headline

On the Trends view, llmprof sums reclaimable spend across all recorded calls and projects it to a month using your observed call rate:

RECLAIMABLE / MO $X · ~N% of spend · projected from M calls/mo

This is the number to act on. It is an estimate (it scales observed local traffic to 30 days), and it is honest about uncertainty: brand-new and hosted open-model prices drift, and findings that cannot be quantified are kept advisory rather than padded into the total.

Where it runs

Detection runs once, off the hot path, when a call is recorded - so the dashboard and the Python SDK share exactly one implementation. SDK traces get the same findings as proxied ones.