Full-text search

whereFullText(fields, query) on the Cascade query builder maps to tsvector on Postgres and $text on MongoDB. Both drivers need an index before the query is cheap — without one, the engine scans every row. This recipe covers the index setup briefly, the Cascade service that runs the search, and the scoring patterns for ranked results.

What whereFullText does

await Post.query().whereFullText(["title", "body"], "cascade orm").get();

• Postgres — translates to a to_tsquery(...) match against the combined to_tsvector(title || ' ' || body). Without a GIN index on that vector expression, it’s a sequential scan.
• MongoDB — translates to $text: { $search: "cascade orm" }. Requires a text index on the relevant fields.

The index lives in the migration; Cascade picks it up automatically once it exists. The Migrations intro covers the column-builder syntax; consult your database’s full-text docs for the specific index expression.

The search service

import { Post } from "../models/post/post.model";

type SearchPostsInput = {
  query: string;
  limit?: number;
};

export async function searchPosts(input: SearchPostsInput) {
  return Post.query()
    .whereFullText(["title", "body"], input.query)
    .orderBy("createdAt", "desc")
    .limit(input.limit ?? 20)
    .get();
}

Plain query builder. The whereFullText call carries the index hint — the rest of the chain is the same vocabulary as every other query.

For combined filters (full-text + structured), chain them together:

return Post.query()
  .where("status", "published")
  .where("authorId", authorId)
  .whereFullText(["title", "body"], input.query)
  .orderBy("createdAt", "desc")
  .limit(input.limit ?? 20)
  .get();

The structured filters narrow the candidate set first; the full-text predicate runs against the remainder.

Ranked results — driver-specific scoring

Default whereFullText results aren’t sorted by relevance — they’re returned in whatever order the index produces, plus your explicit orderBy. For real relevance ranking, project the driver’s score expression with selectRaw and order by it:

import { Post } from "../models/post/post.model";

type RankedSearchInput = {
  query: string;
  limit?: number;
};

export async function rankedSearchPosts(input: RankedSearchInput) {
  return Post.query()
    .whereFullText(["title", "body"], input.query)
    .selectRaw(scoreExpression(input.query))
    .orderByRaw("score DESC")
    .limit(input.limit ?? 20)
    .get<PostRow & { score: number }>();
}

The scoreExpression helper holds the driver-specific bit:

function scoreExpression(query: string) {
  return {
    // Postgres
    expression:
      "ts_rank(to_tsvector('english', title || ' ' || body), to_tsquery('english', ?)) AS score",
    bindings: [query.split(/\s+/).join(" & ")],

    // MongoDB equivalent shape would be { score: { $meta: "textScore" } } via selectDriverProjection
  };
}

For cross-driver code, branch the helper by driver type and emit the right shape per side. See the Expressions guide for the selectRaw / selectDriverProjection mechanics.

Common shapes

“Exact phrase” search

Wrap the query in quotes to ask for an exact phrase rather than individual tokens:

await Post.query().whereFullText(["title"], '"cascade orm"').get();

Both drivers honour the quoting convention.

Excluding stop words

Common words (“the”, “a”, “and”) get filtered by the language analyzer your index was built with. On Postgres, to_tsvector('english', ...) uses the English stop-word list; pick a different language ('french', 'german', …) per the language of your content. On MongoDB, the text index has a default_language field with the same effect.

Stemming

Both drivers stem words by default — “running” and “runs” both match “run”. This is usually what you want; turn it off only if your content has technical terms (product codes, SKUs) where stems would over-match. The migration sets this via the index’s default_language option.

Combining with vector search

Full-text catches exact keyword matches; vector catches semantic similarity. For the best of both, run both queries and merge — covered in detail in the Hybrid search recipe.

Going further

• The expressions vocabulary for selectRaw and orderByRaw — Expressions guide
• Hybrid search combining full-text with vector — Hybrid search recipe
• The full whereFullText signature — Query Builder API reference