💿

DiskANN Index

Keeps hot vectors in RAM, rest on disk · handles billion-scale with small memory

DiskANN keeps hot vectors in fast RAM and stores the rest on disk. Search scans RAM first, then reads targeted disk vectors. Near-HNSW quality with much less memory.

Flow: 📥 Query 🧠 Search RAM 💿 Read Disk 🏆 Top Match
🎯

Query Vector

Editable
🎚️

RAM Cache Size

1 5 3 in RAM
More disk readsFewer disk reads
🧠

RAM Cache

Editable 3 vectors
💿

Disk Storage

Editable 5 vectors
🔍

Click Search to see results

Adjust the RAM slider, edit vectors, then run a search to see how DiskANN works.

💡 Key Idea

RAM as fast guide, disk for full data. 10–100× less RAM than pure HNSW.

✅ Best For

Billion-scale datasets, limited RAM, cost-sensitive production workloads.

Hot vectors stay in RAM, the rest live on disk, letting DiskANN handle datasets far larger than available memory.
It's built for billion-scale search, where keeping everything in RAM (like HNSW) would be too expensive.
The trade-off is slightly higher latency from disk access compared to fully in-memory indexes.
💡
Key Takeaway

DiskANN keeps only the most-used vectors in RAM and the rest on disk, enabling billion-scale search with a small memory footprint.