Section 3 · Data 360 Fundamentals · 20% of exam
3.2 Data 360 Concepts for Agentforce
Exam objective
Describe how grounding and retrieval-augmented generation use Data 360 content to provide relevant context to Agentforce prompts and agents.
Why Agentforce needs data context?
A large language model can generate fluent language without knowing a company’s current policies, product details, or customer-specific records. Grounding supplies relevant context at the time a response is generated. In a Data 360 RAG design, source data is brought into a searchable form, organized into a search index, and queried through a retriever. The retrieved passages are added to a prompt so the model can formulate an answer using the business context. The system does not normally send every source document to the model. It searches for a smaller set of relevant material, which helps keep the prompt focused and within its context limits.
The preparation-to-response path
A useful mental model has two phases. During preparation, data is ingested or selected, mapped, divided into chunks when needed, represented for search, and organized in an index. During a user interaction, the user’s request or a generated search query is sent to a retriever. The retriever searches the configured index, returns relevant results, and the prompt template incorporates those results as context. The model then produces a response from the instructions, the request, and the retrieved evidence. Agentforce can use an action such as Answer Questions with Knowledge to connect this flow to a data library. Each component has a distinct job: source data provides facts, chunking creates manageable passages, indexing makes them searchable, retrieval selects candidates, and generation writes the answer.
Search and relevance
Vector search represents text as numerical embeddings that capture semantic relationships. This can find conceptually similar passages even when the query and source use different words. Keyword search can be useful when exact names, product codes, or phrases matter. Hybrid search combines vector and keyword approaches. The right search design depends on the data and questions: a policy question expressed in varied language may benefit from semantic matching, while an exact identifier may require precise lexical matching. Search is not the same as generation. A search index finds candidate content; the language model uses selected content to produce a response. Improving the prompt cannot repair a source that was never indexed or a retriever that selects irrelevant results.
Data 360 is more than a document folder
Data 360 provides data integration and management capabilities that support this pipeline. A data library automates several setup tasks, while advanced designs can configure search indexes and retrievers individually. Search indexes can work with unstructured content and, in supported configurations, structured content. Data objects and mappings organize source information. Chunk data model objects can represent passages created from text, and index data model objects can represent vectorized searchable content. Students do not need to memorize every object name for a scenario question; they should understand the role of each stage and know where to investigate readiness, configuration, and processing status.
Grounding is not a permission grant
Retrieved context can make an answer more specific, but authorization and governance remain separate. Admins must decide what content should be indexed, what access the agent or user has, and whether an external audience may receive a passage. RAG can also return incomplete or imperfect evidence. Include representative test questions, ambiguous requests, outdated-source cases, and questions with no answer in the source. If the model is expected to decline when evidence is absent, state that behavior in the prompt or agent configuration and test it. When citations are available, expose sources to help users verify answers, while still treating generated output as something to evaluate.
If an answer is inaccurate, trace the full chain: source content, ingestion and indexing status, search configuration, retriever results, prompt context, and response instructions.
Grounding can improve factual relevance but does not make generated responses automatically correct or enforce Salesforce record permissions by itself.
Extra concept: RAG is a pipeline; failures at any stage affect the answer.
Official references:
- What Are Data Libraries? https://help.salesforce.com/s/articleView?id=data_library_concept.htm&language=en_US&type=5
- Enhance Agents with Business Knowledge Using RAG https://trailhead.salesforce.com/content/learn/modules/retrieval-augmented-generation-quick-look/augment-prompts-with-relevant-knowledge
Key takeaways
- Grounding adds relevant business context to prompts at response time.
- Data preparation, indexing, retrieval, and generation are separate stages with separate failure points.
- RAG improves relevance; it does not guarantee correctness or replace authorization and testing.
