1. Agentforce Data Library Question: I have hundreds of Knowledge Articles. But, I need to get responses based on 10 Knowledge Articles alone when using the “Answer Questions with Knowledge” action. How can we achieve this?
Data Categories can be used. When Setting up the Agentforce Data Library, in the Knowledge Settings, use “Filter by Knowledge Data Categories”.
2. How to mask the email address before being exposed to large language models (LLM) in Salesforce?
a. Go to Einstein Trust Layer in the Salesforce Setup.
b. Enable “Large Language Model Data Masking”.
c. In the “Pattern Based” section, under “Sensitive Data”, enable “Email Address”.
3. Have you used ODATA? If yes, what is it?
OData, or Open Data Protocol, is an open standard for building and consuming RESTful APIs. It is based on the principles of REST and web technologies like HTTP, to create queryable and interoperable web services. It provides a way to interact with data by using standard HTTP messages for operations like creating, reading, updating, and deleting (CRUD) data.
4. Retriever in Salesforce Data Cloud
A Retriever is a tool that searches and fetches the most relevant and trusted data from your organization’s connected data sources (structured and unstructured) in Data Cloud to “ground” a Large Language Model’s (LLM) response.
When a user or an AI agent asks a question:
The Retriever takes the query and searches a Search Index (which is essentially a vector database containing your organized, chunked, and vectorized data).
It retrieves the most relevant snippets of data (known as “chunks”).
This retrieved data is then packaged and sent along with the original question to the Large Language Model (LLM).
5. Implicit Sharing
Implicit sharing is a built-in mechanism in the Salesforce platform that automatically grants users access to parent records if they can access related child records and vice versa.
a. Access to a parent account—If you have access to an account’s child case, contact, or opportunity record, you have implicit Read Only access to that account, even with org-wide defaults on Account set to Private.
Note:
To view the parent account via implicit access, you must have access to the related child case, contact, or opportunity record via ownership or through sharing access. You can’t access the parent account if you only have access to the child records through the View All Records or Modify All Records object permissions.
b. Access to child records—Depending on your sharing configuration, if you have access to a parent account, you can have access to the associated child case, contact, and opportunity records. The account owner’s role and settings configured in account sharing rules and manual shares determine the level of access to child records.
https://help.salesforce.com/s/articleView?id=platform.sharing_across_objects.htm&type=5
6. Apex Hammer Test Results
With the Hammer process, Salesforce runs your org’s Apex tests in both the current and new release, and compares the results to identify issues for you.
7. Salesforce Model Temperature
In Salesforce Einstein Studio, Model Temperature can be modified or adjusted.
Higher temperatures mean the model takes more risks by providing more random responses. Lower temperatures mean that generations are less random. Minimum value is 0.0, and maximum value is 1.0
A lower value generates more consistent responses.
8. What is the use of Stop Sequence in Model Configuration
The generated text from the Model is cut at the end of the earliest occurrence of a stop sequence configured. Stop sequences are useful to control the length and the structure of the response.
Example: “\n” to stop the generation with a new line.
For example, if your application generates structured data that should end with an XML-style closing tag like , you can use that tag as a stop sequence. The model will stop as soon as it generates that specific string.
9. What is the runtime lifecycle of an Salesforce Agentforce Action?
Agentforce Actions are how agents get things done.
User Prompt –> Reasoning Engine –> Topic Selection –> Action Execution(Apex Or Flow Or Prompt) –> Response back to the user.
First, the user prompts. For example, “Show My Top 5 Opportunities which need immediate attention”. Salesforce Reasoning Engine will find the right topic. It verifies the scope for the selected topic. Based on the Instructions in the Topic, it will find the matching Action. If the inputs to the Action are available, it will execute the action. Else, it will prompt for the inputs to execute the action. The engine uses the raw data from the action to generate a natural language response using the LLM before displaying it to the User.
10. How does Salesforce Agentforce Agent Instructions override system behaviour?
The Salesforce Reasoning is trained to prioritize specific user-provided instructions over general pre-trained behaviours.
Example:
System Default: Answer any question the user asks.
Agent Instruction: “Never answer questions about competitors. Politely decline. (This overrides the default helpfulness).
11. What is the difference between Tasks, Messages & Turns in Salesforce Agentforce?
Tasks:
Tasks are Agent Actions. A Flow, Apex Class, or Prompt.
Turn:
One cycle of AI reasoning/thought. One user message can trigger multiple reasoning turns to find the right topic.
Message:
A single text bubble (User or Agent) visible to the users.
12. Briefly, provide the Salesforce Agentforce Instruction Hierarchy with an example.
- Agentforce Agent Description, Role and Company
- Topic Instructions
- Agent Action Instructions
13. How does Salesforce Agentforce decide which prompt template to use?
Agentforce Actions can be defined using the Prompt Templates. The Atlas Reasoning Engine finds the right topic. For the right Action(based on multiple Prompt Templates Actions), it uses semantic search suing the Action Name and Description to execute.
14. What is Grounding and how does it prevent hallucinations in Salesforce?
LLM hallucinations are the result of the model attempting to plausibly predict the next word, even when it lacks the answer, relying on generalized patterns learned from its training data (the internet). Grounding prevents this guesswork.
When a user submits a question, with the help of Grounding, the system first intercepts the prompt before sending it to the Large Language Model (LLM). This initial step is a Retrieval Augmented Generation (RAG) process, which involves a Retrieval or search action. The purpose of RAG is to inject the retrieved raw data alongside the user’s prompt. This augmentation minimizes the risk of the LLM generating inaccuracies, often referred to as “Hallucinations.”
15. How does Agent Action Handlers work in Salesforce Apex?
Agent Action Handlers are implemented using standard @InvocableMethod annotations.