Search

The Search engine finds documents by their words in any language, forgiving typos, by the vectors your embedding model makes, or by both. An index keeps its own documents in search.db, or follows a table of an application's database and changes in the same commit as its rows. Searches are exact by default.

const articles = store.index<Article>('articles', { fields: { title: 2, body: 1 }, filters: ['authorId'] })

await articles.set('a-17', { title: 'Как испечь хлеб', body: '…', authorId: 42 })
const { hits, corrected } = await articles.search({ text: 'хлеба', where: { authorId: 42 } })

articles = store.index("articles", Article, fields={"title": 2, "body": 1}, filters=["author_id"])

await articles.set("a-17", Article(title="Как испечь хлеб", body="…", author_id=42))
result = await articles.search(text="хлеба", where={"author_id": 42})

articles, err := search.OpenIndex[Article](store, "articles",
	search.Fields{"title": 2, "body": 1}, search.Filters("author_id"))

err = articles.Set(ctx, "a-17", Article{Title: "Как испечь хлеб", Body: "…", AuthorID: 42})
result, err := articles.Search(ctx, "хлеба", search.Where{"author_id": 42})

let articles = store.index::<Article>("articles").field("title", 2.0).field("body", 1.0).filters(["author_id"]).open()?;

articles.set("a-17", &Article { title: "Как испечь хлеб".into(), body: "…".into(), author_id: 42 })?;
let result = articles.search("хлеба").filter(eq("author_id", &42)).run()?;

What it guarantees

A search by words answers what scoring every document would, by BM25, never an estimate (contract SRCH-01).
A search by a vector answers the exact best by cosine, unless the index asked for recall (contract SRCH-02).
A filter keeps to the documents it passes before ranking (contract SRCH-03).
A document set is found by the next search; a deleted one never again (contract SRCH-04).
A table's index changes in the commit that changes the row (contract SRCH-07).
A misspelt word is read as its nearest, and corrected says so; a word that matches is never changed (contract SRCH-09).

In these guides

Page	Covers
Fields and weights	which text is read, and how much each field counts
Typos	the nearest word, `corrected`, `typoTolerance`
Filters	`where`, lists, ranges, `search.or`
Vectors	dimensions, exact search, `recall` and clusters
Words and vectors together	`vectorWeight`, `rerank`
A table's index	`db.searchIndex`, triggers, `ready`
Rebuilds	changing fields and filters while the index answers

New to Search? Finding old messages in the Handbook indexes a chat's messages, and How search finds explains the index.

Reference

Bun and Node · Python · Go · Rust

#What it guarantees

#In these guides

#Reference

What it guarantees

In these guides

Reference