Skip to main content

Relevance Examples

Start with representative queries and expected results. Apply one tuning change at a time, rerun the same cases, and keep the smallest rule set that works.

Add collection synonyms​

Treat laptop, notebook, and portable computer as equivalent product terms:

curl --fail-with-body -X POST \
  http://localhost:9200/collections/products/synonyms/portable-computers \
  -H "Content-Type: application/json" \
  -d '{"synonyms":["laptop","notebook","portable computer"]}'

List and remove the rule:

curl http://localhost:9200/collections/products/synonyms
curl -X DELETE \
  http://localhost:9200/collections/products/synonyms/portable-computers

Prefer collection synonyms unless the vocabulary truly applies everywhere.

Remove low-value words​

curl --fail-with-body -X POST \
  http://localhost:9200/collections/products/stopwords \
  -H "Content-Type: application/json" \
  -d '{"word":"the"}'

Do not add domain terms merely because they occur frequently; doing so can make valid queries impossible. See Stopwords for local and global scope behavior.

Pin a result for an exact query​

curl --fail-with-body -X POST \
  http://localhost:9200/collections/products/overrides/featured-keyboard \
  -H "Content-Type: application/json" \
  -d '{
    "rule": {"query":"keyboard","match":"exact"},
    "includes": [{"id":"keyboard-2","position":1}]
  }'

To suppress a known bad result, use "excludes":["document-id"]. Overrides are best for explicit merchandising or editorial decisions, not broad relevance tuning.

Save reusable search parameters​

curl --fail-with-body -X PUT \
  http://localhost:9200/presets/storefront-search \
  -H "Content-Type: application/json" \
  -d '{
    "query_by": "title,description,tags",
    "facet_by": "category,brand",
    "highlight_full_fields": "title",
    "limit": 24
  }'

curl http://localhost:9200/presets/storefront-search

Presets centralize shared defaults. Keep user input such as q, pagination, and allowed filters separate from the preset unless the application requires fixed values.

Boost important query terms​

curl --fail-with-body -X POST \
  http://localhost:9200/collections/products/documents/search \
  -H "Content-Type: application/json" \
  -d '{
    "q": "wireless^2 keyboard",
    "query_by": "title,description,tags",
    "limit": 10
  }'

Use moderate boosts and verify that they improve more than one hand-picked query. For custom ranker configuration, see Ranking.

A simple evaluation loop​

Keep a small table in version control:

[
  {"q":"travel keyboard","expected_top_ids":["keyboard-1"]},
  {"q":"gaming keyboard","expected_top_ids":["keyboard-2"]},
  {"q":"notebook accessories","expected_top_ids":["stand-1","sleeve-1"]}
]

Run each query before and after a change, compare the first few IDs, and record latency as well as ranking quality. This catches synonym and override changes that fix one query while degrading another.