mirror of
https://github.com/headroomlabs-ai/headroom.git
synced 2026-11-22 13:29:19 +08:00
fix(examples): drop the item-cap claim from the comparison demo
The 15-item setting applies only to lossy row selection, not to lossless compaction or passthrough, so the demo no longer presents it as a cap. The simulation calls its figure a rough 5x estimate. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
90370d2cdd
commit
0ff4cbf340
@@ -453,9 +453,7 @@ def run_simulation():
|
||||
|
||||
print(f"\n Total tool output: {total_tool_tokens:,} tokens")
|
||||
print(f" With 3 iterations, baseline input would be: ~{total_tool_tokens * 2:,} tokens")
|
||||
print(
|
||||
f" With Headroom (default cap ~15 items), estimated: ~{total_tool_tokens // 5:,} tokens"
|
||||
)
|
||||
print(f" With Headroom (rough 5x estimate), estimated: ~{total_tool_tokens // 5:,} tokens")
|
||||
print(
|
||||
f" Estimated savings: ~{total_tool_tokens * 2 - total_tool_tokens // 5:,} tokens (~80%)"
|
||||
)
|
||||
@@ -497,8 +495,7 @@ def print_summary(baseline_runs: list[AgentRun], headroom_runs: list[AgentRun]):
|
||||
Headroom reduced input tokens by {pct_saved:.1f}% across all scenarios.
|
||||
|
||||
Key optimizations applied:
|
||||
- SmartCrusher: Compressed tool outputs from 50-200 items to the most relevant
|
||||
ones (the default router keeps about 15)
|
||||
- SmartCrusher: Compressed the large JSON tool outputs (50-200 items each)
|
||||
- Context preserved: Agent still found the right information
|
||||
|
||||
This translates to:
|
||||
|
||||
Reference in New Issue
Block a user