IGNIS

Appearance

Changelog - 2026-07-08 ​

New Feature Breaking Change

In one line. The Typesense connector now supports vector/semantic and hybrid search, federated and union multi-search across collections, and declarative synonyms - reachable through one unified search({ mode }) method.

What changed ​

  • Unified search({ mode }). One repository method discriminated by mode: keyword (full-text), semantic (vector), hybrid (blended), and raw (native passthrough) - replacing many specialized methods with a single, typed, OpenAPI-generated schema.
  • Vector fields, both embedding modes. field.vector('embedding', ...) in the collection DSL. Supply dimensions + distance for client-provided vectors, or embed: { from, model } to have Typesense generate the vector at index time - no external embedding pipeline required.
  • Cross-collection multiSearch. dataSource.multiSearch({ searches, union?, commonParams?, options? }) - federated by default (side-by-side results per collection), or merged into one ranked set with union: true.
  • Declarative synonyms. synonyms: [{ id, synonyms, root? }] on defineSearchCollection, provisioned automatically at configure(). Multi-way synonyms (no root) make every listed term interchangeable; one-way synonyms (root set) expand a root query to its synonyms.
  • defineSearchController factory. Generates a POST /search and POST /multi-search endpoint from a repository, mirroring the existing CRUD controller factory - each endpoint is customizable or can be disabled.
  • Capabilities model. getCapabilities() now reports search: { vector, multi, union, synonyms }, so calling code can probe what the connected engine supports at runtime.
  • Extended search parameters. Faceting, highlighting, grouping, and result-tuning fields are now accepted as optional camelCase fields on every non-raw mode, mapped to Typesense's snake_case wire names automatically.
  • Raw escape hatches, everywhere. search({ mode: 'raw', params }) for native single-collection params, getConnector().multiSearch(...) for native multi-search, and getClient() for the underlying Typesense client - any engine feature the typed API doesn't model is still reachable.

Who is affected ​

  • Anyone using Typesense search. Vector/semantic search, multi-search, synonyms, and the controller factory are all new and opt-in - no action needed to keep existing keyword search working.
  • Anyone who adopted the raw search({ params }) passthrough introduced on 2026-07-05. The call now requires mode: 'raw' - see Breaking changes below.
  • Everyone else. No action needed.

Breaking changes ​

1. Raw search now requires mode: 'raw' ​

search() is a single method discriminated by a required mode field, so a bare search({ params }) call (the shape introduced on 2026-07-05) no longer type-checks.

Before

typescript
const result = await articleRepository.search({ params: { q: 'typescript', query_by: ['title'] } });

After

typescript
const result = await articleRepository.search({ mode: 'raw', params: { q: 'typescript', query_by: ['title'] } });

Details ​

  • No engine lock-in in the base layer. All new verbs sit on neutral contracts (ISearchConnector, ISearchQueryDialect, the search() mode union) with engine-agnostic shapes, so a second search engine can slot in later without touching src/base.
  • Vector search and prefix matching don't mix. semantic and hybrid default to prefix: false, since prefix matching is meaningless for vector search and remote embedders (OpenAI, Google, ...) reject it outright. Callers can still override via the prefix field.
  • Synonyms use the v30+ global synonym-sets model - one Typesense synonym set per collection, linked at configure(). Runtime management is available via the connector (upsertSynonymSet / getSynonymSet / listSynonymSets / deleteSynonymSet / linkSynonymSets).
  • No sync guarantees. No retry, circuit-breaker, or change-data-capture is built into the framework - keeping the database and the search index in sync remains caller policy.