Class: SearchClient

The AI Search client. Construct it with createSearchClient (which reads a resolved config) or directly for one-off use. All reads resolve the OBO workspace client from the active execution context.

Constructors

Constructor

new SearchClient(config?, workspaceClientFactory?, readBackend?): SearchClient

Parameters

config?

ResolvedSearchConfig = ...

workspaceClientFactory?

() => WorkspaceClient

readBackend?

SearchReadBackend

Returns

SearchClient

Methods

addDocuments()

addDocuments(reference, documents, signal?): Promise<{ count: number; index: string; }>

Add or update documents in a direct-access index.

Parameters

reference

string

documents

Record<string, unknown>[]

signal?

AbortSignal

Returns

Promise<{ count: number; index: string; }>

createIndex()

createIndex(name, options?): Promise<IndexInfo>

Create an AI Search index with as little ceremony as possible. Two shapes:

Delta Sync (the default): pass sourceTable; Databricks computes embeddings from the text column and keeps the index synced. The embedding model is resolved automatically when not named.
Direct Access: omit sourceTable and pass embeddingDimension; you write vectors yourself via addDocuments.

Everything else infers: the endpoint from the plugin config, the primary key (id), the text column (text / content / body), and the vector column (embedding). Returns the created index’s IndexInfo.

Parameters

name

string

options?

CreateIndexOptions = {}

Returns

Promise<IndexInfo>

deleteDocuments()

deleteDocuments(reference, ids, signal?): Promise<{ count: number; index: string; }>

Delete documents from a direct-access index by primary key.

Parameters

reference

string

ids

(string | number)[]

signal?

AbortSignal

Returns

Promise<{ count: number; index: string; }>

deleteIndex()

deleteIndex(reference, signal?): Promise<void>

Delete an index.

Parameters

reference

string

signal?

AbortSignal

Returns

Promise<void>

embed()

embed(texts, model?, signal?): Promise<number[][]>

Embed text via a Databricks embedding serving endpoint, returning one vector per input. Used to seed a direct-access index and to turn a search query into a query vector - Databricks only manages embeddings for Delta Sync indexes, so a direct-access index (no Delta table, no warehouse) needs the client to embed on write and on query. The endpoint is resolved the same way as for index creation when not named.

Parameters

texts

string[]

model?

string

signal?

AbortSignal

Returns

Promise<number[][]>

ensureEndpoint()

ensureEndpoint(name?, options?): Promise<void>

Ensure a Vector Search endpoint exists, creating a STANDARD one when it does not. Optionally wait for it to come online. Idempotent.

Parameters

name?

string

options?

EnsureEndpointOptions = {}

Returns

Promise<void>

ensureIndex()

ensureIndex(name, options?): Promise<IndexInfo>

Create the index if it does not already exist, otherwise return the existing one. Idempotent - safe to call on every boot to guarantee an index is present.

Parameters

name

string

options?

CreateIndexOptions = {}

Returns

Promise<IndexInfo>

getIndex()

getIndex(reference, signal?): Promise<IndexInfo>

Fetch an index’s live definition.

Parameters

reference

string

signal?

AbortSignal

Returns

Promise<IndexInfo>

index()

index(reference): SearchIndex

A handle bound to one index (by full UC name or configured alias).

Parameters

reference

string

Returns

SearchIndex

listIndexes()

listIndexes(endpoint?, signal?): Promise<string[]>

List the indexes hosted on a Vector Search endpoint (name + type only).

Parameters

endpoint?

string

signal?

AbortSignal

Returns

Promise<string[]>

provision()

provision(name, options?): Promise<IndexInfo>

Ensure an index exists, is online, and (optionally) holds seed data - the “wire up a real index on boot” path. Idempotent and cheap to call every boot: it creates the endpoint and index only if missing, waits for them to come online, and seeds documents ONLY when the index is still empty.

For the demo/dummy-data case this needs no Delta table and no warehouse: the default is a MANAGED direct-access index (Databricks embeds the text column on write and query), so seeding is just an addDocuments of plain rows and search-by-text works immediately.

Parameters

name

string

options?

ProvisionOptions = {}

Returns

Promise<IndexInfo>

resolveEmbeddingModel()

resolveEmbeddingModel(requested?, signal?): Promise<string | null>

Resolve an embedding endpoint id for creating a Delta Sync index. Reuses the model resolver: a configured / passed name is fuzzy-matched against the live catalogue, otherwise the highest-ranked embedding endpoint is chosen.

Parameters

requested?

string

signal?

AbortSignal

Returns

Promise<string | null>

search()

search(query, options?): Promise<{ count: number; hits: object[]; index?: string; query: string; }>

Search one index. index may be a full UC name, a configured alias, or omitted to use the default index. Returns hits sorted most-relevant-first.

Parameters

query

string

options?

SearchOptions & object = {}

Returns

Promise<{ count: number; hits: object[]; index?: string; query: string; }>

syncIndex()

syncIndex(reference, signal?): Promise<void>

Trigger a sync of a Delta Sync index from its source table.

Parameters

reference

string

signal?

AbortSignal

Returns

Promise<void>

universalSearch()

universalSearch(query, options?): Promise<{ count: number; hits: object[]; index?: string; query: string; }>

Fan a query across several indexes and merge the hits, sorted by score - the “universal search” a single box over many collections needs. Each index is searched concurrently; an index that errors is logged and skipped so one bad index does not sink the whole search.

Parameters

query

string

options?

UniversalSearchOptions = {}

Returns

Promise<{ count: number; hits: object[]; index?: string; query: string; }>