TBNilles/LEANN
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

benchmarks: fix and extend HNSW+DiskANN recompute vs no-recompute; docs: add fresh numbers and DiskANN notes

Browse Source
This commit is contained in:
Andy Lee
2025-08-14 12:18:07 -07:00
parent 79ca32e87b
commit b13b52e78c
4 changed files with 94 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
docs/configuration-guide.md
View File

@@ -363,12 +363,23 @@ Trade-offs:
Real-world quick benchmark (HNSW, 5k texts; script `benchmarks/benchmark_no_recompute.py`):
```text
recompute=True: ~6.58s; size ~1.1MB
recompute=False: ~0.10s; size ~16.6MB
recompute=True: ~7.55s; size ~1.1MB
recompute=False: ~0.11s; size ~16.6MB
Conclusion: no-recompute is much faster but uses more storage; recompute is smaller but has higher first-hop latency.
```
DiskANN (5k texts; same script, final rerank strategy):
```text
build(recompute=False): size ~24.8MB
build(recompute=True, partition): size ~5.7MB
search recompute=False: ~0.250s (on recompute-build)
search recompute=True (final rerank): ~0.120s (on recompute-build)
Conclusion: DiskANN's recompute-build enables partitioning to reduce storage; enabling final rerank further improves latency while keeping traversal PQ-fast.
```
## Further Reading