Choose quality. Understand the token trade-off.

LLM match uses an LLM to judge whether extracted field values match the expected values. F1 balances finding expected results with avoiding incorrect ones; generation rubrics use a 1-to-7 scale.

Document token expense is relative to GPT-6.1 Sol with Schema only (1.00×), calculated from measured input/output token usage and the evaluation's rates. It excludes service processing, embeddings, training and other charges; it isn't a current price quote.

Create an analyzer with a schema alone, or add documents to its knowledge base and label the expected answers. This evaluation compares Schema only and With 5 labeled examples.

Task names follow the public field methods and segmentation terminology. Speech measures fields from audio, not transcription. Graphs show GA evaluation averages.

Video compares nine models across different execution cohorts, not a controlled model-only experiment or every historical model. Headline Generation F1 is a separate aggregate across the two video datasets. Video whiskers show mean ± one standard deviation (SD), in percentage points, not confidence intervals or proof of a statistically significant difference. No video token expense, cost or latency is paired with these scores.