Curiosity

Full Text Search

Full-text (BM25) search is the default retrieval path. Configuration boils down to three questions: which fields are indexed, how they rank, and what counts as "the same" word.

1. Pick the searchable fields

  1. Settings → Search → Full Text Search.
  2. + Add more — pick (node type, field) pairs.
  3. Toggle Searchable on the node type to control whether matches surface in the default workspace search results.

Recommended fields per type:

Type Typical fields Notes
SupportCase Summary, Content Boost Summary higher; Content is longer-tailed.
KbArticle Title, Body, Tags Title and Tags should outweigh Body.
Product Name, Description, Sku Sku boosted heavily — exact matches are intent signals.

Avoid indexing high-cardinality opaque fields (UUIDs, hashes) — they bloat the index without helping recall.

Each indexed field has its own settings, reachable from the index's advanced settings and each explained by its (?) tooltip. The ones that change what matches:

Setting What it does
Searchable Off leaves the field out of users' searches; field-scoped queries from custom code still reach it.
Index Aliases On (the default) also indexes the node's aliases whose language matches this index, so a node is found by its other names. Off restricts matching to the field's own text.
Index Lemma Also indexes each word's base form, so running matches run. Costs roughly one extra term per word.
Index Stop Words Keeps the, and, … in the index. Grows it, but lets a query made only of stop words match.
Fuzziness How many letter edits are tolerated: 0 exact only, 1 or 2 a fixed edit distance, -1 automatic by word length.
Normalize Diacritics / Ignore Case Match Muller against Müller, and match case-insensitively. Both only fully apply after a reindex.

2. Adjust ranking with field boosts

Each indexed field has a boost value (the + / − controls). Matches in higher-boost fields score higher. Useful defaults:

  • Title-like fields: 3–5×
  • Summary fields: 2×
  • Body fields: 1× (baseline)
  • Tags / labels: 2–3×

Adjust iteratively against the evaluation framework — don't tune by gut.

3. Synonyms

Domain abbreviations and internal jargon hurt recall when they aren't declared. Add them under Settings → Search → Synonyms, one entry per term, sorting the related words into Synonyms, Alternates or Blocked.

Two things to keep in mind: an entry is one-directional unless you turn on its link toggle (mbp → MacBook Pro does not imply the reverse), and synonyms apply when a query runs, so a change needs no re-index.

See Synonyms for the editor, the language scoping, and the API.

Multi-language behavior

Each indexed field is tied to its node's language detection (via the NLP pipeline). The search engine stems, lower-cases, and removes stop-words per-language. For mixed-language corpora, no extra setup is needed — each document is processed with its detected language.

From code

The same configuration drives SearchRequest from custom endpoints:

var req = SearchRequest.For("battery drain");
req.BeforeTypesFacet = new HashSet<string> { N.SupportCase.Type };

// Optional: limit to a subset of indexed fields for this query.
req.RestrictedFields = new[] { "Summary" };

var query = await Graph.CreateSearchAsUserAsync(req, CurrentUser, CancellationToken);
return query.Take(20).Emit();

Setting RestrictedFields on a request narrows ranking to a slice of the configured index — useful when you want a "title-only" search lane from the same index.

Validating changes

After changing field boosts or synonyms:

  1. Run the evaluation framework eval set. Look for regressions in recall@k and MRR.
  2. Spot-check a handful of representative queries from production logs.
  3. If you've added a new indexed field, rebuild the index: Settings → Search Index → Rebuild. A synonym change needs no rebuild.
© 2026 Curiosity. All rights reserved.