Optimizing AI Workflows: What I Learned from Four Text-Analysis Trials

Описание и применение

Оптимизация AI‑рабочих процессов через разделение задач, улучшение системных сообщений и удаление лишнего текста сократили токены и стоимость на 16 %.

Задача

Complex AI workflows become expensive due to repeated small design decisions, leading to high token usage and cost.

Результат

Improved system messages and caching cut uncached input cost by ~73% and overall cost by ~16%. Removing unnecessary natural‑language output reduced output tokens by ~15% and saved ~$2.8 per run. A single mis‑instruction caused repeated function calls, highlighting that cache stats alone don’t guarantee efficiency.

Рабочий процесс

• Extract sentences with GPT 5.6 Sol (low reasoning)
• Tokenize sentences with GPT 5.4 mini
• Extract multi‑word terms with GPT 5.6 Terra (medium reasoning)
• Batch token classifications (5 tokens per batch) using Terra
• Reuse previous outputs to reduce decision space
• Record configuration, timestamps, token usage, and context for each model call
• Apply caching and remove unnecessary final assistant messages

Доказательства

Between the two TEXT 1 trials, the estimated cost of uncached input fell by almost 73%.

Источник

The output cost also fell by almost 15%.

Источник

Removing the single full stop instruction caused the model to generate explanatory text, increasing output from 234,000 to 426,000 tokens.

Источник