Improve search quality

Search quality refers to the quality of search results in terms of ranking and recall as perceived by the user.

Ranking is the order of items, and recall is the number of relevant items retrieved. An item (or document) is any piece of digital content Google Cloud Search can index, such as Microsoft Office documents, PDF files, database rows, or unique URLs. An item consists of:

• Structured metadata
• Indexable content
• ACLs

Cloud Search uses various signals to retrieve and rank results. You can influence these signals through settings in the schema, item content and metadata, and the search application. This document helps you improve search quality by modifying these signal influencers.

Caution: Because search quality is subjective, you must understand business use cases and user expectations before tuning settings. Undisciplined tuning can result in a poor search experience.

For a summary of settings, see the Summary of search quality settings.

Note: Before reading this document, see Create and register a schema for an explanation of schema structure.

Influence topicality score

Topicality is the relevance of a search result to the query terms. Cloud Search calculates topicality based on:

• The importance of each query term.
• The number of hits (how often a query term appears in content or metadata).
• The match types between query terms and indexed items.

To influence a text property's topicality score, define the RetrievalImportance in your schema. A match on a property with high RetrievalImportance results in a higher score.

For example, if you store software bug history, users likely query by bug name. Set the RetrievalImportance for the name to HIGHEST. For descriptions, use DEFAULT.

{
  "objectDefinitions": [
    {
      "name": "issues",
      "propertyDefinitions": [
        {
          "name": "summary",
          "textPropertyOptions": {
            "retrievalImportance": { "importance": "HIGHEST" }
          }
        },
        {
          "name": "description",
          "textPropertyOptions": {
            "retrievalImportance": { "importance": "DEFAULT" }
          }
        }
      ]
    }
  ]
}

For HTML documents, tags like <title> and <h1> help determine term importance. If ContentFormat is TEXT, ItemContent uses DEFAULT importance. If it is HTML, importance depends on HTML properties.

Influence freshness

Freshness measures how recently an item was modified, using createTime and updateTime in ItemMetadata. Older items rank lower.

You can influence freshness by adjusting freshnessProperty and freshnessDuration in FreshnessOptions. The freshnessProperty lets you use a custom date or timestamp instead of updateTime. For bugs, you might use a due date so items with dates closest to now rank higher.

{
  "objectDefinitions": [
    {
      "name": "issues",
      "options": {
        "freshnessOptions": { "freshnessProperty": "duedate" }
      }
    }
  ]
}

Use freshnessDuration to specify when an item is considered out-of-date. A high value, like 100 years, nullifies the freshness effect. This is useful for static data like employee profiles. For news, use a small value like three days.

Influence quality

Quality measures an item's accuracy and usefulness. Specify a value between 0 and 1 using SearchQualityMetadata. Higher values boost ranking. For example, you might boost benefits documents authored by HR.

{
  "metadata": {
    "searchQualityMetadata": { "quality": 1.0 }
  }
}

Influence using field type

You can influence ranking based on enum or integer properties using OrderedRanking:

• NO_ORDER (default): No effect on ranking.
• ASCENDING: Higher values boost ranking.
• DESCENDING: Lower values boost ranking.

For bug priority, use DESCENDING if HIGH is 1 and LOW is 3. For a votes property, use ASCENDING.

Influence ranking through query expansion

Query expansion uses synonyms and spelling to improve results.

Use synonyms

Cloud Search uses public web content for synonyms. You can also define custom synonyms for organization-specific terminology. See Define synonyms.

Use spelling

Cloud Search provides spelling suggestions in SpellResult. It also uses spell corrections as synonyms to find documents.

Influence ranking through Search Application settings

A Search Application provides contextual information. You can influence ranking using the following settings.

Adjust the scoring configuration

Specify a ScoringConfig to disable freshness or personalization. Disabling personalization removes owner and interaction boosts. See Customize the search experience.

Adjust the source configuration

Specify data source-level settings:

• Source importance: Items from a HIGH importance source rank higher.
• Crowding: Limits results from a single source (default is 3) to ensure diversity.

Personalization

Personalization prioritizes results based on the user:

• Item ownership: Boosts items owned by the user.
• Item interaction: Boosts items the user recently viewed, edited, or commented on.
• User clicks: Boosts items the user previously clicked. Don't modify the result URL, as it is used to collect this data.

Query interpretation

The query interpretation feature converts user queries into structured, operator-based queries using schema definitions. This provides precise results even with minimal keywords. See Structure a schema for optimal query interpretation.

Item language

Cloud Search demotes items whose language doesn't match the query language. Set languageCode in RequestOptions to the user's preference.

Item context

Use contextAttributes to increase ranking for items relevant to the query context, such as location or department.

Item popularity

Cloud Search boosts items that users frequently click in recent queries. Don't modify search result URLs, as they are used to track these clicks.

Summary of search quality settings

The following tables provide a summary of search quality settings:

Schema settings

Setting	Location	Recommendation	Details
ItemContent	ItemContent	Recommended	Topicality
RetrievalImportance	RetrievalImportance	Recommended	Topicality
FreshnessOptions	FreshnessOptions	Optional	Freshness

Indexing settings

Setting	Location	Recommendation	Details
createTime/updateTime	ItemMetadata	Recommended	Freshness
contentLanguage	ItemMetadata	Recommended	Item language
owners	ItemAcl	Recommended	Item ownership
Custom synonyms	_dictionaryEntry	Recommended	Use synonyms
quality	SearchQualityMetadata	Optional	Influence quality
Interaction data	interaction	Optional	Item interaction
OrderedRanking	OrderedRanking	Optional	Field type

Search application settings

Setting	Location	Recommendation	Details
Disable personalization	ScoringConfig	Recommended	Personalization
Source importance	SourceCrowdingConfig	Optional	Source configuration
Crowding	SourceCrowdingConfig	Optional	Source configuration

Next Steps

1. Structure a schema for optimal query interpretation.
2. Define synonyms.