Complete workflow comparison
18–27%Fewer tool calls
3.4–4.7%Less returned text
30 synthetic runs, a separate experiment Method
MEASURED, NOT ASSUMED
Efficiency, with evidence.
Less repeated reading. Essential facts retained. Every number has a checkable basis.
73.6%
Less input in the largest task packet
Compared with reading the full source corpus
676 / 676Required facts retained
This compares input material, not total billing savings or model quality.
Method and limitations +
This measures input material, not total billing savings or model accuracy. Full JSON includes extra metadata. The baseline is a full-source read, not optimized search. 150 synthetic tasks · 39 active tasks · o200k_base · 5,000-token packet budget
Cost of adoption
Apache 2.0 open sourceUse, modify and integrate under the license
No model call to compileLocal source and context processing
Actual costs still depend on models, caching, output and rework.
