RetrievalSummaryDefinition Metadata in SFDX

Architectural Overview#

RetrievalSummaryDefinition represents the declarative blueprint for Retrieval-Augmented Generation (RAG) grounding within the Salesforce Einstein 1 and Agentforce architectures. In enterprise generative AI implementations, Large Language Models (LLMs) must be grounded with proprietary enterprise data (such as product documentation, knowledge articles, support transcripts, or compliance policies) to produce accurate, hallucination-free responses.

Rather than injecting massive unstructured documents directly into LLM context windows—which inflates token consumption and degrades reasoning latency—RetrievalSummaryDefinition choreographs a two-stage pipeline:

1. Semantic Vector Search: Retrieves the most contextually relevant document chunks from Data Cloud Vector Indexes or hybrid search retrievers.
2. Contextual Distillation & Summarization: Applies a summarization model to distill retrieved passages into concise, grounded context before passing them to Prompt Builder or Agentforce autonomous agents.

Directory and File Conventions#

RetrievalSummaryDefinition artifacts are stored in the retrievalSummaryDefinitions directory within the default package folder.

Key Metadata Fields#

The definition establishes data connector bindings, vector search index links, and summarization parameters.

Declarative XML and Package Manifest Examples#

The following example establishes a retrieval summary pipeline grounding customer service agents on knowledge base troubleshooting guides:

<?xml version="1.0" encoding="UTF-8"?>
<RetrievalSummaryDefinition xmlns="http://soap.sforce.com/2006/04/metadata">
    <masterLabel>Knowledge Support Retriever</masterLabel>
    <developerName>Knowledge_Support_Retriever</developerName>
    <dataConnector>Support_Docs_Vector_Index</dataConnector>
    <searchType>Hybrid</searchType>
    <summaryModel>sfdc_einstein_gpt_summarizer</summaryModel>
    <maxResults>5</maxResults>
    <status>Active</status>
    <description>Retrieves top 5 support knowledge articles and generates a synthesized grounding summary for Case copilot prompts.</description>
</RetrievalSummaryDefinition>

Package Manifest (package.xml)#

<?xml version="1.0" encoding="UTF-8"?>
<Package xmlns="http://soap.sforce.com/2006/04/metadata">
    <types>
        <members>Knowledge_Support_Retriever</members>
        <name>RetrievalSummaryDefinition</name>
    </types>
    <version>68.0</version>
</Package>

CLI Retrieval and Deployment Commands#

Retrieve and deploy RetrievalSummaryDefinition metadata with the Salesforce CLI:

# Retrieve all retrieval summary definitions
sf project retrieve start --metadata RetrievalSummaryDefinition

# Deploy definition to target sandbox
sf project deploy start --source-dir force-app/main/default/retrievalSummaryDefinitions/Knowledge_Support_Retriever.retrievalSummaryDefinition-meta.xml

# Validate definition deployment
sf project deploy validate --metadata RetrievalSummaryDefinition:Knowledge_Support_Retriever

Architect Best Practices and Gotchas#

• Data Cloud Index Dependencies: The vector search index referenced in dataConnector must be provisioned and active in Data Cloud before this metadata type can be successfully deployed.
• Token Window Management: While maxResults controls passage volume, ensure the intermediate summaryModel does not exceed prompt size limits when combined with user conversation histories.
• Security and Masking: Grounding pipelines respect Einstein Trust Layer data masking and zero-data retention policies. Verify that data ingested into the underlying vector index complies with data residency regulations.

Frequently asked questions

What does RetrievalSummaryDefinition do in Salesforce AI?

It defines retrieval-augmented generation pipeline rules, extracting relevant passages from vector search indexes and summarizing them to ground generative prompts.

Where is RetrievalSummaryDefinition stored in an SFDX project?

It is located in the retrievalSummaryDefinitions directory with the .retrievalSummaryDefinition-meta.xml file extension.

How does RetrievalSummaryDefinition connect with Data Cloud?

It links to Data Cloud vector search indexes and search retrievers, executing semantic search queries against unstructured documents and knowledge bases.

Where are these retrieval summaries consumed?

They are referenced inside Prompt Builder prompt templates, Agentforce agents, and Einstein copilot reasoning engines.