Section 3 · Data 360 Fundamentals · 20% of exam
3.6 Choosing a Grounding Approach for a Scenario
Exam objective
Select an appropriate grounding approach by matching the business question to the data type, source, freshness, search behavior, and response requirements.
Start with the information need
A strong grounding design begins with the question the agent must answer. Identify the source of truth, whether the information is structured or unstructured, how often it changes, what the user is allowed to see, and how the answer should be presented. A question about a published policy may fit a Knowledge article or approved document. A question requiring a current CRM field or transaction value may need direct record context or an appropriate data action rather than a document search. A request for recent public information may need Search the Web rather than an internal data library. The right design follows the data and task; using RAG for every question can add unnecessary complexity.
Match the source to the design
Use a Knowledge-based library when authoritative Salesforce Knowledge articles contain the information and the required fields can be indexed. Use an uploaded-file library when approved content is maintained in documents and those files are the appropriate source. Use a custom retriever when the data is available through a suitable Data 360 search configuration or when a supported advanced retrieval design is required. New web-search data libraries are not supported, so consider the separate Search the Web action or web retriever in Prompt Builder for current external information where allowed. If the library does not support the required source, choose another grounding method.
Consider freshness and precision
For stable policies, an indexed library can provide useful passages, provided the organization maintains updates and confirms they are processed. For rapidly changing account or order information, an agent may need a real-time record lookup or action that queries the current value. For exact identifiers, ensure search can match the identifier; for natural-language questions, semantic or hybrid search may help. If the answer depends on several source collections, determine whether supported multiple-library or ensemble-retriever patterns fit. The standard library limits each feature to one assigned library at a time, so source architecture matters. Consider where context is assembled, not only where data is stored.
Account for access, evidence, and risk
Determine whether the answer may contain personal, confidential, or customer-specific data. Identify whose permissions govern the retrieval and action, whether the content is appropriate for external users, and whether the response must include citations or human review. A retriever does not grant a user permission to see records, and grounding is not a substitute for data governance. If the answer drives a consequential action, combine the model with deterministic checks and an appropriate approval step. For informational responses, citations or visible sources can help users check the evidence.
Work through an exam scenario
Suppose a customer asks, “Can I return this item, and what steps do I follow?” First identify whether eligibility depends on a static policy or live purchase details. The policy passages may come from Knowledge or approved files; the order date and status may require a live record lookup. The agent may need both types of context through separately designed capabilities, even though one library is assigned to a feature at a time. It should retrieve the policy, verify the order details through an authorized action, apply deterministic eligibility rules where required, and explain the result with source evidence. Test valid, expired, missing-order, and ambiguous-item cases. Use this pattern elsewhere: identify facts, choose a source, retrieve or query it, enforce rules, and verify the response.
For scenarios, identify the fact needed, trusted source, freshness, access, and response checks.
Do not choose an Agentforce Data Library for every data need. Real-time records, deterministic calculations, and current external web information can require different capabilities.
Extra concept: Combine retrieval with deterministic actions when both policy context and exact decisions are needed.
Official references:
- Choose a Data Source for Your Library https://help.salesforce.com/s/articleView?id=ai.data_library_choose_data_source.htm&language=en_US&type=5
- Ground Agentforce in Data https://help.salesforce.com/s/articleView?id=ai.agent_parent_data.htm&language=en_US&type=5
- Use the Answer Questions with Knowledge Action https://help.salesforce.com/s/articleView?id=ai.agent_setup_data_sources.htm&language=en_US&type=5
Key takeaways
- Match the grounding method to the source, freshness, search, access, and response requirements.
- Use indexed knowledge or files for stable content; use direct actions for current record facts or deterministic rules.
- Test source coverage, access boundaries, missing evidence, citations, and edge cases before release.
