Backendless, browser-based full-text search at tens-of-millions scale: Summa 2 updated (now supports vector search)

Internet services IPFS
1

New methods such as vector queries have been added.
sparse vectors (BMP, MaxScore, Seismic), dense search (flat, TQ, IVF-TQ, binary IVF, ScaNN), and hybrid ranking

The early 2023 implementation used Tantivy to compile to WASM; as of 2026, it uses a self-developed version.

It is also very AI-friendly now. You can try asking AI to help you create a search for literary works or novels. It seems to have also fixed a bug in Chinese word segmentation.
However, using the tokenizer built into JS for segmentation may be better.

const text = "今天天气很好，我准备去公园散步。";

const segmenter = new Intl.Segmenter("zh-CN", {
  granularity: "word",
});

The WASM size is only 9 MiB. If you don’t need vector queries, you can ask AI to help remove features including IpfsIndex, sparse vectors, dense search, LocalIndex, and others. The size can be reduced to 3 MiB and .

Those interested can look into it.

Example

IPFS Search (old version)

repo 20.36T
dois 12.74T

2 Likes