Developer cuts GPT-4 job-listing pipeline cost 63% after fixing rate limits and batch logic

A developer built a production system that scores over 10,000 job listings daily using GPT-4 function calling, vector search, and a REST API. Early runs were costly and slow, with the first full-day batch taking 47 minutes and costing $86, while aggressive retry logic later caused a three-hour delay. Switching from single-chunk to multi-chunk extraction with focused schemas reduced structured-output errors from 12% to under 2%, at the cost of more API calls per listing. Choosing pgvector over Pinecone and OpenAI's smaller embedding model over the larger one cut monthly embedding costs from roughly $1,872 to $144. Adopting OpenAI's Batch API, which offers 50% off in exchange for deferred processing, brought the per-run cost down from $86 to $32, a 63% reduction.