LLM usage
What the transform pipeline spends on the LLM and where it goes. The pipeline runs two to four times a day, so every figure here is a sum over a day's runs rather than a single run — a row read as a day would understate spend by however many runs it dropped.
Cost per day
A day's spend is the sum of its transform runs. The dashed line is the window mean.
Cache hit rate
The number that explains the cost above. A run with a cold cache costs several times what the same work costs warm, so a dip here and a spike there are the same event.
Cost by call site
Each day's spend split by the stage that made the call. This is what says whether a rise came from more listings needing the LLM or from one stage losing its cache.
- ask-llm-with-results
- ask-llm-to-categorise
- ask-llm
- identify-multiple-movies
- identify-shorts
Call sites over the window
Totals across every run collected. Cost per thousand calls is the figure to compare stages by — a stage making many cheap cached calls is not the one to look at first, and a per-call figure would round to zero for all of them.
| Call site | Calls | Cache misses | Miss rate | Cost | Per 1k calls |
|---|---|---|---|---|---|
| ask-llm-with-results | 23,004 | 12,740 | 55.4% | $8.66 | $0.376 |
| ask-llm-to-categorise | 34,170 | 13,262 | 38.8% | $1.87 | $0.055 |
| ask-llm | 44,557 | 17,891 | 40.2% | $1.69 | $0.038 |
| identify-multiple-movies | 3,327 | 1,639 | 49.3% | $0.814 | $0.245 |
| identify-shorts | 7,297 | 3,314 | 45.4% | $0.456 | $0.062 |
By month
One row per month in the window. Cost per day is the fair comparison between a finished month and one still running.
| Month | Days | Runs | Calls | Cached | Prompt tokens | Cost | Per day |
|---|---|---|---|---|---|---|---|
| 2026-08 | 14 | 33 | 60,057 | 57.0% | 63.2M | $7.22 | $0.516 |
| 2026-09 | 13 | 29 | 52,298 | 56.0% | 54.4M | $6.26 | $0.482 |
Biggest prompts
Only the single largest prompt per run is logged, so this is a tally of which venues turn up as the worst offender — not a ranking of every venue's prompt size. A venue here is one whose listings page is large enough to be worth trimming before it reaches the model.
| Venue | Largest prompt | Runs topped | Call sites |
|---|---|---|---|
| goethe.de | 57.9k chars | 21 | ask-llm-with-results |
| bfi.org.uk-southbank | 52.6k chars | 40 | ask-llm-with-results |
Venue coverage
How many of the venues in a transform run needed the LLM at all. The rest were handled by the deterministic parsers.
On the most recent run, 190 of 414 venues needed the LLM — 46% of the pipeline.