contextweaver

Auto-generated by make trend. Do not edit by hand. Source: benchmarks/results/history/*.json (one snapshot per release).

Release-over-release view of the deterministic benchmark metrics. Latency is excluded — it is environment-dependent and not comparable across release machines. This page is visibility only; PR-time regression gating lives in benchmarks/gating.yaml + scripts/benchmark_gate.py (#491).

Releases recorded: 2 (0.16.0 … 0.17.0).

Routing recall@k by catalog size

release	size=50	size=83	size=1000
0.16.0	0.5649	0.3825	0.1475
0.17.0	0.5649	0.3825	0.1475

Routing MRR by catalog size

release	size=50	size=83	size=1000
0.16.0	0.4978	0.3242	0.1456
0.17.0	0.4978	0.3242	0.1456

Routing precision@k by catalog size

release	size=50	size=83	size=1000
0.16.0	0.1191	0.0800	0.0310
0.17.0	0.1191	0.0800	0.0310

Context pipeline quality

release	mean token reduction	items dropped	dedup removed
0.16.0	64.31%	7	4
0.17.0	65.01%	11	4

---

Capturing a release snapshot

make benchmark   # refresh benchmarks/results/latest.json
python scripts/render_trend.py --snapshot <version> \
    --from benchmarks/results/latest.json
make trend       # re-render benchmarks/trend.md