TRADE-OFFS
What we're honest about
94% is against full-file reads
Claude Code already uses grep and partial reads. Real-world savings vs normal behavior are lower. Full-file is our reproducible baseline.
Claude Code already uses grep and partial reads. Real-world savings vs normal behavior are lower. Full-file is our reproducible baseline.
Monorepos dilute recall
On Go's fiber (396 files), recall dropped to 0.07@10. One-feature-per-file repos hit R=1.00. Focused files retrieve best.
On Go's fiber (396 files), recall dropped to 0.07@10. One-feature-per-file repos hit R=1.00. Focused files retrieve best.
Embedding model matters
bge-small-en-v1.5 (384d) is fast, not SOTA. Bigger models lift recall but add latency. We chose speed; <1s re-index at 96% cache.
bge-small-en-v1.5 (384d) is fast, not SOTA. Bigger models lift recall but add latency. We chose speed; <1s re-index at 96% cache.
What actually worked
Simple heuristics over ML. SQLite over specialized DB.
Simple heuristics over ML. SQLite over specialized DB.