You are here:
Hybrid Search Best Practices
Enhance your hybrid search retrieval and search relevance by chunking the correct fields, selecting pre-filter fields, and adding ranking factors.
Use chunking fields that provide context and semantic meaning. You can select multiple fields for chunking your data. Consider how you expect the retrieved results to be used in your gen AI application. For example, select fields such as title, body, or description of your records. Fields that have a small set of values, such as Open or Closed, aren’t ideal for chunking.
Increase the precision of search results by adding filter fields to narrow the results. Use filters that make sense for your application. For example, pre-filtering on the case status field can improve the relevance of search results on case conversation transcripts.
Refine result relevance with ranking models by sending relevance-specific parameters. Use advanced formulas from the Reciprocal Rank Fusion (RRF) and Linear Fusion Ranker (LFR) models to tweak parameter values and achieve relevant results.
Enhance relevance by adding ranking factors to your search index configuration.
- Add the popularity of records as a factor, such as selecting fields that track ViewCount. For instance, when you index a DMO created from CRM knowledge articles, import the entity KnowledgeArticleViewStat into Data 360, and use KnowledgeArticleViewStat.ViewCount as the popularity factor.
- Add recency of records as a factor. For example, when you index a DMO created from CRM knowledge articles, use LastModifiedDate__C field as the recency factor.
{
"label": "record9357",
"developerName": "record9357",
"description": "A sample description",
"sourceDmoDeveloperName": "Knowledge_kav_Home__dlm",
"chunkDmoName": "chunk9357",
"vectorDmoName": "vector9357",
"chunkDmoDeveloperName": "chunk9357",
"vectorDmoDeveloperName": "vector9357",
"searchType": "HYBRID"
"vectorEmbedding": {
"vectorEmbeddingRelatedFields": [
{
"relatedDmoDeveloperName": "Knowledge_kav_Home__dlm",
"relatedDmoFieldDeveloperName": "ssot__Name__c",
"relationships": [
{
"sourceDmoDeveloperName": "Knowledge_kav_Home__dlm",
"sourceDmoFieldDeveloperName": "Content_c__c",
"targetDmoDeveloperName": "Knowledge_kav_Home__dlm",
"targetDmoFieldDeveloperName": "Content_c__c"
}
]
}
]
},
"rankingConfigurations": [
{
"relatedDmoDeveloperName": "Knowledge_kav_Home__dlm",
"relatedDmoFieldDeveloperName": "LastModifiedDate__c",
"rankType": "RECENCY",
"developerName": "rank_time_freshness",
"relationships": [
{
"sourceDmoDeveloperName": "Knowledge_kav_Home__dlm",
"sourceDmoFieldDeveloperName": "Content_c__c",
"targetDmoDeveloperName": "Knowledge_kav_Home__dlm",
"targetDmoFieldDeveloperName": "Content_c__c"
}
]
}
],
"chunkingConfiguration": {
"fieldLevelConfigurations": [
{
"sourceDmoDeveloperName": "Knowledge_kav_Home__dlm",
"sourceDmoFieldDeveloperName": "Content_c__c",
"config": {
"id": "passage_extraction",
"userValues": [
{
"id": "max_tokens",
"value": "512"
},
{
"id": "strip_html",
"value": "true"
}
]
}
}
]
},
"vectorEmbeddingConfiguration": {
"embeddingModel": {
"id": "e5_large_v2",
"userValues": [
{
"id": "dimension",
"value": "1024"
},
{
"id": "max_token_limit",
"value": "512"
}
]
},
"index": {
"id": "HNSW",
"userValues": [
{
"id": "hnswEfConstruction",
"value": "8"
},
{
"id": "M",
"value": "4"
}
