Prompts contain long, unused context. What is the most direct saving?
Why B is the answer
Remove what the task does not need
Fewer tokens lower cost and latency.
In detail
- Unused context costs tokens on each call.
- Removing it is the most direct saving.
- Measure quality after trimming.
Use case
An illustrative scenario, not a real incident.
Each prompt carries pages of unused context. Removing it cuts tokens on every call, at once.
See it in different ways
stateDiagram-v2 [*] --> CheckRetrieval CheckRetrieval --> TuneChunks: wrong chunks CheckRetrieval --> TunePrompt: right chunks, poor answers TunePrompt --> FineTune: style still not met TuneChunks --> [*] FineTune --> [*]
flowchart LR S0["Find unused context"] --> S1["Remove it"] --> S2["Check quality"] --> S3["Measure savings"]
Key point: Unused context costs money on each call.
Why the other options are wrong
- A. Increase the deployment's TPM
- More TPM adds capacity, not savings.
- C. Move the model to a different region
- Region does not cut tokens.
- D. Add a second retrieval step on each call
- A second retrieval step adds cost.