turbopuffer

RSS: https://turbopuffer.com/blog/rss.xml

基于对象存储的向量与全文检索：快速、便宜 10 倍、极具可扩展性。

安息吧，向量数据库

turbopuffer 团队宣布将存储架构从"向量为主键"的 v1/v2 设计重构为 v3：ANN 索引降级为二级索引，主键改为独立于向量聚类的结构，从而同时改善文本、正则、聚合与向量搜索性能，并减少存储/写放大。文章详细梳理了 v1 的 SPANN/SPFresh 层次聚类索引、v2 引入的属性过滤与 BM25 全文检索如何都围绕 ANN 地址组织，并指出该设计导致多向量文档存储重复、SPFresh 再平衡时级联搬移动整个文档及其倒排索引、以及查询引擎向量化块大小被聚类限制在 100–200 条等问题。 v3 已实现 CI 全通过，后续将逐步优化性能并公开基准测试。turbopuffer 当前承载超 1T 文档、每秒 10M+ 写入、25k+ 查询，客户包括 Cursor 与 Notion。

点赞

收藏1 天前

再见了，向量数据库

<p>Turbopuffer 正在推动搜索技术的前沿发展。为此，我们必须从根本上重新设计存储架构，使向量索引不再是主要存储方式。</p>

点赞

收藏3 天前

为什么把 embedding 放进 turbopuffer 内部能降低搜索延迟

turbopuffer 官方博客讨论为何将 embedding 计算内置到向量数据库可显著降低搜索延迟。作者称原生 embedding 不只是便利功能，还能为搜索查询省掉数百毫秒的等待时间——通常这部分耗时来自客户端或中间层把文本转成向量再发给数据库的网络往返。对于自建向量检索管线或比较 embedding 服务架构的工程师来说，这篇短帖提供了明确的性能动机，也隐含了“把模型塞进存储层”这一设计取舍，值得点开看具体数据。

点赞

收藏4 天前

用后训练开源小模型做大规模代码搜索

<p>Applied Compute 与 turbopuffer 携手证明，经过微调以使用预计算检索索引的小型专用检索模型，能够在成本和延迟仅为前沿模型极小一部分的情况下，实现令人印象深刻的检索效果。</p>

点赞

收藏29 天前

命名空间分支功能：即时克隆与独立读写

turbopuffer 推出 Namespace Branching 功能，通过 branch_from 实现命名空间的即时克隆，基于 copy-on-write 机制，克隆时间与命名空间大小无关。分支完全独立，支持多级分支链，无子分支数量限制。每次分支费用固定 0.032 美元，存储按逻辑字节计费。适用场景包括代码库索引、测试流水线、开发环境隔离和快照查询。

点赞

收藏46 天前

Building turbopuffer

点赞

收藏61 天前

RAG is dead, right??

点赞

收藏116 天前

Training SID-1 to beat GPT-5 at search with 1k+ QPS RL

SID-1 is an agentic search model that is 24x faster than GPT-5.1-high, 374x cheaper than Sonnet 4.5, and achieves 1.9x higher recall than traditional RAG pipelines. Here's how we trained it using larg...

点赞

收藏135 天前

Simon Eskildsen on scaling Shopify, building turbopuffer, and the future of databases

点赞

收藏142 天前

Mixing numeric attributes into text search for better first-stage relevance

turbopuffer now allows you to combine attribute values into the scoring function of text queries. Ranking by attribute helps achieve better relevance in the first stage with the same scalability chara...

点赞

收藏158 天前

Building the database for trillion-scale AI search

点赞

收藏174 天前

Retrieval After RAG: Hybrid Search, Agents, and Database Design

点赞

收藏205 天前

Object storage-native database for search

点赞

收藏208 天前

Rust zero-cost abstractions vs. SIMD

A customer query was taking over 4× longer than it should have. The profiler pointed at Rust code we'd assumed was free. We followed the trail all the way down to assembly to find the true cost.

点赞

收藏213 天前

A distributed queue in a single JSON file on object storage

点赞

收藏224 天前

ANN v3: 200ms p99 query latency over 100B vectors

Our latest ANN release supports scales of 100+ billion vectors in a single search index, with 200ms p99 query latency at 1k QPS and 92% recall.

点赞

收藏254 天前

Designing inverted indexes in a KV-store on object storage

How we redesigned our inverted index structure using fixed-sized posting blocks to achieve 10x smaller indexes and dramatically better throughput.

点赞

收藏261 天前

Why BM25 queries with more terms can be faster (and other scaling surprises)

I analyzed how BM25 query latencies scale with document count and top_k. Longer queries scale less efficiently, and essential terms impact performance in some surprising ways.

点赞

收藏268 天前

Vectorized MAXSCORE over WAND, especially for long LLM-generated queries

turbopuffer has improved text search performance up to 20x by adopting the same text search algorithm as Apache Lucene, a vectorized variant of block-max MAXSCORE

点赞

收藏291 天前

FTS v2: up to 20x faster full-text search

turbopuffer's full-text search engine is getting a major upgrade for 20x better full-text search performance

点赞

收藏303 天前

Faster vector search

点赞

收藏324 天前

Billion-scale vector storage for RAG

点赞

收藏333 天前