Salesforce Agentforce-Specialist Practice Exam Questions & Answers

To ensure that content generated by a large language model (LLM) is appropriately screened for toxic language, the Agentforce Specialist should create a Trust Layer audit report within Data Cloud. By using the toxicity detector type filter, the report can display toxic responses along with their respective toxicity scores, allowing Universal Containers to monitor and manage any toxic content generated with a high level of confidence.

Option C is correct because it enables visibility into toxic language detection within the Trust Layer and allows for auditing responses for toxicity.

Option A suggests checking a toxicity detection log, but Salesforce provides more comprehensive options via the audit report.

Option B involves creating a flow, which is unnecessary for toxicity detection monitoring.

Salesforce Trust Layer Documentation: https://help.salesforce.com/s/articleView?id=sf.einstein_trust_layer_audit.htm

The correct answer is C. In a multilingual RAG design, the embedding model must support semantic similarity across the languages involved in both the source content and user queries. Since the articles are in Spanish and users may ask questions in French or English, selecting the multilingual-e5-large embedding model broadly for the search index is the correct approach. Option A and Option B are too narrow because each excludes one of the languages in the complete requirement. The retriever does not simply match keywords; it compares vector representations of the query and indexed chunks. A multilingual embedding model improves cross-language retrieval by placing semantically similar text from different languages closer in vector space. Salesforce Data 360 search index guidance identifies multilingual content and embedding-model selection as key retrieval considerations.

Einstein Trust Layer'sToxicity Scoringcategorizes content on a scale of0 to 1, where1 indicates the highest level of toxicity(e.g., harmful, biased, or inappropriate language). This scoring helps organizations filter unsafe AI-generated content. A score of 1 triggers mitigation actions, such as blocking the response or alerting administrators.

A score of 0would indicate no toxicity (B is incorrect).

The scoring system does not use 'least toxic' as a category (A is misleading).

Salesforce Help Article:Einstein Trust Layer -- Toxicity Scoring('Interpreting Toxicity Scores' section).

Einstein GPT Safety Overview: 'Mitigating Harmful Content with Toxicity Detection.'

In Salesforce, when creating a prompt template for the sales team, you can include data from related objects such as Opportunities that are linked to an Account. The best method to ground the AI model and provide relevant information from related records, like Opportunities, is by using merge fields.

Merge fields in Salesforce allow you to dynamically reference data from a record or related records, like Opportunities for a given Account. In this scenario, the Agentforce Specialist needs to pull data from the default related list of Opportunities associated with the Account. This is achieved by using merge fields, which pull in data from the standard relationship Salesforce creates between Accounts and Opportunities.

Option A (referencing a custom related list) and Option C (using formula fields with Einstein-related lists) do not align with the standard, practical grounding method for this task. Custom lists would require additional configurations not typically necessary for a basic use case, and formula fields are typically not used to directly fetch related list data for prompt generation in templates. The standard and straightforward method is using merge fields tied to the default related list of opportunities.

Salesforce Reference:

Merge Fields in Templates: https://help.salesforce.com/s/articleView?id=000387601&type=1

Grounding Data in Prompts: https://developer.salesforce.com/docs/atlas.en-us.salesforce_ai.meta/salesforce_ai/grounding_data_prompts

The correct answer is A because Retriever Playground is the correct no-code tool for validating retrieval behavior before production deployment. It allows the specialist to test individual retrievers, run representative queries, inspect retrieved chunks, and tune retrieval parameters or filters. This is different from Agentforce Testing Center, which validates agent behavior across utterances, subagents, and actions; it is not the most direct tool for testing whether a retriever surfaces the right chunks. Data 360 Query Editor is also not the right choice because the requirement is retriever validation, not manual querying. For RAG quality, the specialist must verify the retrieval layer before evaluating the final generated response. Retriever Playground is built specifically for that retriever-level validation workflow.

The most critical technical consideration when pairing a prompt template's grounding data with a chosen Large Language Model (LLM) is the relationship between the two. The correct action is to review the model limitation in Prompt Builder versus the grounding data size (C).

Every LLM has a fixed context window limit, typically expressed in tokens (the model's units for processing text). This token limit defines the maximum amount of input data (the prompt template text + all the dynamic grounding data) and output data the model can handle in a single request.

The grounding data, which is pulled dynamically from Salesforce records (e.g., related lists, long text fields, Flow outputs), varies significantly in size from one record to the next. If the combined size of the prompt and the dynamic data for a specific record exceeds the LLM's token limit, the generative AI request will fail with a 'token limit exceeded' error. The Agentforce Specialist must proactively design the template to limit the amount of data retrieved (e.g., using Flow to summarize related lists or querying only essential fields) to ensure it stays within the chosen model's capacity.

Option A is incorrect because the Einstein Trust Layer's token limit primarily relates to PII masking and is a security-related capacity, not the fundamental model's context window. Option B is incorrect because OFFSET is a SOQL query function used for pagination, which is irrelevant to ensuring the total size of the final assembled prompt (template + data) fits within the model's token limit.

Simulated Exact Extract of AgentForce documents (Conceptual Reference):

'A major challenge in prompt template design is managing the Large Language Model (LLM) token limit against the volume of grounding data. The specialist must always Review the model limitation in Prompt Builder versus the grounding data size before activation. LLM context windows (token limits) are fixed per model, but dynamic prompt components---such as merge fields from related lists or long text area fields---can cause the total size of the prompt to vary significantly by record. To prevent random token limit failures, the prompt instructions and grounding logic (Flow/Apex) must be explicitly constrained to retrieve only the essential data required to answer the query, ensuring the combined input stays well below the LLM's defined capacity.'

Simulated Reference: AgentForce Prompt Builder Best Practices Guide, Section 4: Performance and Scalability, p. 92.

The correct answer is A. Credential-Based User Verification ties the service interaction to the authenticated Experience Cloud user, so record access and record changes are evaluated through that verified user's security context. That explains why audit fields such as Last Modified By show the Community User rather than the generic Agent User. This is not an error; it is the expected governance outcome when the agent is operating with verified user identity. Option B is incorrect because system-context flow execution would explain bypassed sharing behavior, not why the authenticated user appears in audit fields. Option C is incorrect because the scenario explicitly states Credential-Based User Verification, not token-based verification. For regulated environments, this behavior is valuable because audit trails show which verified customer identity caused the record update.