We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Lobsters·13d ago·5 min read
Two years of vector search at Notion: 10x scale, 1/10th cost
TL;DR
Notion scaled its vector search infrastructure by 10x while reducing costs by 90% over two years, enhancing its ability to retrieve semantically relevant content.
✦ Why It Matters
Evaluate your current search infrastructure for potential cost and performance optimizations using vector search techniques.
Key Takeaways
- Implement semantic embeddings for improved search relevance.
- Consider serverless architectures to reduce operational costs.
- Utilize change detection to optimize data processing.
- Explore new technologies like Ray for cost-effective embeddings.
How It Works
Notion's vector search uses semantic embeddings to represent text as points in high-dimensional space, allowing for meaning-based retrieval. The architecture evolved from a dual-path indexing system to a serverless model, decoupling storage from compute, which significantly reduced costs and improved scalability.
Change detection was implemented using hashes to identify modifications in page content, allowing for targeted updates rather than full re-uploads.