Use-case preset
Recruiting candidate screening cost calculator
Parse resume and JD into a fit score with reasoning; batch workload.
Each prompt contains a full resume (1–3 pages) plus a job description, totaling 4–8k tokens. Output is a score, a fit summary, and a list of qualification gaps — typically 300–600 tokens. The 85/15 ratio reflects the document-heavy input. Context is set to 16k to handle senior-role JDs that include detailed technical requirements alongside longer resumes.
Batch processing is appropriate; hiring managers check results at the start of their shift, not in real time. Cached prompt percent is 30–40% because the JD repeats across all candidates for a given role, while the resume portion is unique each time. The main accuracy risk is false negatives — set up a calibration set with human-labeled decisions and measure recall before deploying. At scale, a small model fine-tuned on your labeled data often outperforms a large general model on this structured scoring task.