turbopuffer is a serverless vector and full-text search database built natively on object storage designed to cut costs at scale. [1, 2]
Core Architecture
- Object Storage First: Data is stored primarily on cloud object storage (like
- or
- ), drastically lowering storage costs compared to traditional in-memory vector databases.
- Smart Caching Layer: Uses RAM and NVMe SSD caching in front of object storage to keep hot data fast, achieving sub-10ms latencies for frequent queries.
- Serverless Scaling: Automatically handles indexing, caching, and storage tiering without requiring cluster management or server bin packing. [2, 6]
Key Features
- Hybrid Search: Supports vector search, BM25 full-text search, and hybrid queries combining both.
- Native Metadata Filtering: Allows precise filtering alongside vector similarity matching.
- High Throughput: Handles massive write volumes and scales namespaces independently past hundreds of millions of documents. [2, 4, 7]
Notable Use Cases
- Notion: Powers production embeddings for Notion AI search, handling billions of text chunks while cutting millions from their database bill.
- Cursor: Leverages the object-storage architecture for fast RAG operations across massive codebases, reducing costs significantly. [6, 8]
AI responses may include mistakes.
No comments:
Post a Comment