Agentforce Specialist Data 360 explained
Data 360 Fundamentals is worth 20% of the Salesforce Agentforce Specialist exam — around 12 of the 60 scored questions. It is the shortest section in the outline and the one candidates most often treat as background reading.
It is not background. It is the machinery that makes grounding work, and when an agent gives a confidently wrong answer, this is almost always where the fault lies.
Salesforce names just two objectives, and both matter.
The Agentforce Data Library
A curated body of content an agent or prompt template can draw on for grounding. The considerations the outline asks about are the practical ones:
- What goes in. A focused, relevant library produces better answers than everything you have. More content is not better content.
- How it is maintained. Content changes, and a library that is not refreshed quietly goes stale.
- Who can access it, and that grounding should not surface content a user could not otherwise see.
- What it is for — unstructured content that would not fit as record data.
The three concepts
Chunking, indexing and retrievers. Be able to explain each in one sentence, and say what goes wrong when each is configured badly.
Chunking
Splitting content into pieces small enough to be retrieved precisely.
- Too large and a chunk spans several topics, so its representation matches none of them well and retrieval becomes imprecise.
- Too small and a chunk loses the context that makes it meaningful on its own.
- Boundaries matter. Splitting that respects document structure — sections, headings — beats splitting at arbitrary character counts, because a passage cut in half is a passage lost.
Chunking is the highest-leverage decision in the whole grounding pipeline, and the most common cause of an agent citing the wrong part of a long document.
Indexing
Making chunks searchable. Content is processed into a form that supports finding it by meaning rather than exact wording.
The practical consequence: an index must be maintained. New content is not available until it is indexed, and an index that lags behind its source produces confidently outdated answers.
Retrievers
The component that takes a query and returns the most relevant chunks to ground the response.
- How many chunks are returned — too few starves the answer, too many dilute it.
- Relevance: whether what came back actually addresses the question.
- What happens when nothing relevant is found. An agent that invents an answer rather than admitting the gap is a design failure, not a model failure.
The diagnostic that answers most questions
When grounding produces a bad answer, ask in order:
- Is the content in the library at all? If not, no amount of prompt work helps.
- Is it indexed and current? Stale index, stale answer.
- Did the retriever return the right chunks? If not, look at chunking and the query.
- Only then consider the prompt or the model.
Steps one to three are the answer far more often than step four, and the exam reflects that.
Sample questions
Question 1. An agent grounded on a Data Library answers fluently but consistently cites an adjacent, irrelevant section of long reference documents. What should be examined first?
- A. The Trust Layer configuration
- B. The chunking strategy used before indexing
- C. Model access management settings
- D. The agent’s channel configuration
Show answer
Answer: B
Fluent answers drawn from the wrong part of a document point at how content was split before indexing, since oversized or badly bounded chunks cause imprecise matching. The Trust Layer governs security and privacy, model access controls which model runs, and channel configuration concerns where the agent is surfaced.
Question 2. New policy documents were added to a Data Library this morning, but the agent still answers using the previous versions. What is the most likely explanation?
- A. The prompt template needs to be reactivated
- B. The Trust Layer is masking the new content
- C. The new content has not yet been indexed
- D. The agent is connected to the wrong channel
Show answer
Answer: C
Content must be indexed before a retriever can find it, so newly added documents are unavailable until indexing completes. Activation applies to prompt templates rather than library content, the Trust Layer does not withhold updated content, and channel configuration is unrelated.
Question 3. A team asks whether adding every document the company owns to a Data Library will improve answer quality. What is the best response?
- A. No — a focused, relevant library retrieves more precisely than an indiscriminate one
- B. Yes — more content always improves grounding
- C. Yes, provided the chunk size is increased proportionally
- D. It makes no difference, since the retriever filters everything anyway
Show answer
Answer: A
A focused, relevant library improves retrieval precision, whereas indiscriminate content increases the chance of retrieving something plausible but irrelevant. Indexing cost and storage are secondary concerns, and answer quality does not scale with volume.
What to practise
Build one Data Library from a handful of documents, ground an agent on it, and ask five questions you know the answers to. Then change the chunking and ask the same five again.
The difference in answer quality — with the model, the prompt and the agent all unchanged — is the entire argument for why this 20% section exists.