Example: Agentic RAG with Advanced Data 360 Setup

Approaches to Building RAG-Powered Agents

There are two main approaches to setting up a RAG-powered agent.

Scenario in this Example

RAG is expert at extracting relevant information from unstructured data, which includes documents, web sites, and free-form text fields in structured data. This scenario uses free-form text (subject and description) that you add to a Salesforce Case object. An agent can use this information to inform user questions about product issues. Other data source options are described elsewhere in Salesforce Help and Trailhead.

Step 1: Set Up a Salesforce Developer Edition Org

This step walks you through instructions to get and prepare a Salesforce Developer Edition org, which is required to complete the remaining steps in this example.

To sign up, click https://www.salesforce.com/form/developer-signup/?d=pb.

Note

• If you’d prefer to use a different Salesforce org, verify that it offers comparable functionality: Data 360 is enabled and provisioned, Einstein is enabled, Agentforce is enabled, Prompt Builder is available, and it’s OK for you to add cases. Note that your user permissions and other org settings can affect your ability to complete all the steps in this example.
• If you're following along with the instructions and your org runs into unexpected issues, try refreshing your browser window. You can also try going back and repeating previous instructions.

1. Log into the Salesforce Developer Edition org.
2. Check your Data Cloud setup. From Setup, in the Quick Find box, enter Data Cloud Setup. If you see Your Data Cloud instance is live under Your Home Org Details, then Data Cloud is already enabled.

   Data Cloud turned on and running

   
3. If Data Cloud isn’t already enabled, to enable it, follow the instructions in-app or in Turn On Data Cloud in Data Cloud Help). Click Get Started to initiate the setup process.

   Get Started with Data Cloud Setup

   

   

   Note A Data Cloud initial setup can take around 30 minutes or more.
4. When finished, verify that Data Cloud is enabled. Confirm that Your Data Cloud instance is live appears under Your Home Org Details.
5. Check your Einstein Generative AI Setup. From Setup, in the Quick Find box, enter Einstein Setup. Make sure that Einstein, Global Language Support for Prompt Builder, and Deploy Prompt Templates are all turned on.

   Turn On Einstein Options

   

   

   Note When turned on, it can take a few minutes for Einstein enablement to take effect.
6. From Setup, in the Quick Find box, enter Agent, and then select Agentforce Agents. To enable Agentforce, turn on Agentforce.

   If you can't find Agentforce Agents, refresh your browser window or wait a little longer while Einstein enablement takes effect.

   Turn On Agentforce

   

Step 2: Add Case Data

Create at least five Case records that name a product and describe a problem that a customer is having.

1. In App Launcher, enter Service and open Service Cloud.
2. Select Cases -> New and add your cases. For each case, specify a subject, a description that details a customer problem, and select a case origin. Skip other fields and accept the defaults. Here’s some example data to enter.

   Example Subject	Example Description
   Power cord caught fire	The power cord caught fire the first time I tried to use my toaster oven to cook a pizza.
   Toaster oven won’t start	The toaster oven is plugged in but when I push the power button, nothing happens. Nothing lights up. There’s no power.
   Toaster temperature not heating up enough	I turned the toaster oven to the broil setting but the temperature doesn’t go above lukewarm even after 10 minutes.
   Toast burns on lowest temperature setting	The toaster oven always burns my toast, even when I select the Light setting.
   Crumbs caught fire	Crumbs from my bagel dropped to the bottom of the toaster oven and caught fire.
   Toaster turns off suddenly for no reason	If I’m cooking anything for more than 15 minutes, the toaster oven shuts off automatically.
   Toaster oven arrived broken	The glass window was broken when I opened the package.
   Package arrived crushed	The toaster oven was visibly damaged when I opened the crushed delivery package.

Step 3: Ingest Case Data into Data Cloud​

RAG-powered solutions use search indexes that are created in Data Cloud. To ingest your free-form case data into Data Cloud, create a data stream.

1. In App Launcher, enter and select Data Cloud.
2. Select the Data Streams tab, click New, select Salesforce CRM, and click Next.
3. Click View Objects, search for and select Case, and click Next.

   Select the Case Object

   
4. For the Object Category, select Profile.
5. All fields are selected by default. In the Standard Fields tab, select only these fields (deselect the others).

   Field	Description
   System Modstamp	Required
   Subject	A summary of the customer’s issue that can be used to answer the customer’s question.
   Description	Additional details that can be used to answer the customer’s question.
   Case ID	Required but not needed for this simple use case. For more advanced use cases, you can use the Case ID to get additional related information, such as from an Account.

   

   Note The Case Comment field can also provide useful data for RAG searches, but it’s not needed for this simple example.
6. Accept the selections in the Custom Fields tab. Then click Next.
7. Click Deploy.
   

   Note Ingesting data can take time.
8. To check status, in the Data Streams menu, click the down arrow next to your new data stream, and select Update Status.

When Last Run Status is Success and Total Records has a number, your data is in Data Cloud.

Data Ingested into Data Cloud

Step 4: Inspect Your Data Lake Object (DLO)​

The data stream creates a Data Lake Object (DLO) called Case_Home__dll. It’s useful to inspect the ingested data, which can help you debug and optimize your RAG solution.

1. Go to the Data Explorer tab in Data Cloud, select Data Lake Object as the Object Type, and select Case_Home. Don’t change the data space.
2. Click Edit Columns. Move Potential Liability to Available Columns if it’s not already there. Move Subject to Selected Columns if it’s not already there. Leave all other columns where they are, and click Done.

   Edit Columns for Case_Home DLO

   
3. Click the Description column twice to sort and show your case data at the top.

   Case Data in Case_Home DLO

   

Tip Developers: To query your Case DLO, click Copy SOQL and copy the query to use in Developer Console or in the Data Cloud Query Editor. You can also query the DLO using a REST client like Postman.

Step 5: Map Your Case DLO to a Case DMO​

Map a DLO to a Data Model Object (DMO) to add an additional layer of structure to your data and make it easier to query and interact with. Map the Case_Home DLO to the Case DMO.

Note DLOs and DMOs are powerful tools for building RAG solutions. For example, you can perform data cleansing on your DLO, map multiple data sources to a single DMO, or perform identity resolution.

1. In the Data Lake Objects tab, open Case_Home.
2. In the Data Mapping panel, click Start.
3. Under Data Model Entities, click Select Objects, search for Case in the All Objects panel, click the + sign next to it, and click Done. DLO and DMO fields are automatically mapped.

   DLO-to-DMO Mapping

   
4. Save your work.

Step 6: Inspect Your Data Model Object (DMO)​

The mapping performed in the previous step created a DMO called ssot__Case__dlm.

1. In Data Explorer, for Object, select Data Model Object, then search for and select Case.
2. Click the Description column twice to sort and show your case data at the top.

   Data in the DMO

   

Tip Developers: To query your Case DMO, click Copy SOQL and copy the query to use in Developer Console or the Data Cloud Query Editor. You can also query the DLO using a REST client like Postman.

Step 7: Build Your Search Index​

A search index makes your text searchable so that your RAG solution can find and retrieve chunks of text that are relevant to the user’s input. In this example, you create a hybrid search index to find relevant content based on both semantic similarity and keywords.

1. In Data Cloud, select the Search Index tab, and click New.
2. Select Advanced Setup and click Next.
3. Select Hybrid Search as the Search Type and Case as the Source Object, then click Next.
4. Click Manage Fields and select Description and Subject, then click Save.
5. Click Next repeatedly to accept the remaining default values, and then click Save to create your search index.
6. Click the Search Index tab to refresh it. Your Case search index is ready when the Search Index Last Run Status is Ready (about 10 minutes).
7. Click the Data Model tab to see the created search index files:

   Search Index Files

   

   • text chunks (Case_chunk_dlm)
   • a hybrid index (Case_index_dlm)

It’s helpful to know how to inspect text chunks for debugging purposes or feature optimization. For example, chunks might reveal why an LLM is responding in unexpected ways. In Prompt Builder, you’re able to view the retrieved chunks, as you’ll see later in this example.

Tip Developers:

• To inspect the data chunks in the search index you created, or to ask natural language query questions, you can use the query API in Apex or the queryDataV2 endpoint in the Data Cloud Connect REST API.
• To see the configurations for all search indexes in an org, use /ssot/search-index.
• Searching with APIs helps you build a proof of concept solution and debug. You can also try different types of search, such as prefiltering or using a vector search, without building a retriever each time.

Step 8: Build a Retriever

A retriever is a wrapper for Einstein Search operations. You create a custom retriever in AI Models to refine the search so that the retrieved content is more relevant to your use case.

Note An individual retriever can fetch data from only one search index. A search index can index only one DMO. If your feature needs to ground on data from multiple data sources, use an ensemble retriever.

1. In Data Cloud, select AI Models , select Retrievers, and click New Retriever.
2. Select Individual Retriever and click Next.
3. Select Data Cloud and accept the default data space.
4. Select the DMO (Case) and search index configuration (Case Hybrid Search), then click Next.
5. For filters, accept the default (All Documents) and click Next.
6. For Fields to Return, add these three fields:
   1. Case object’s Subject field (DirectAttribute > Case > Subject)
   2. Case object’s Description field (DirectAttribute > Case > Description)
   3. Chunk field (RelatedAttributes > Case chunk > Chunk)
   

   Tip Select the field name first. The associated field label is selected automatically.

   Fields to Return in the Retriever

   
7. Click Next, then click Save. Name this retriever and create it.
8. On the Retrieval Datails page, copy the API Name, which you'll use in the next step.
9. Click Activate.

Step 9 - Create a Prompt Template and Add Your Retriever​

A prompt template defines the structure of the prompt that is sent to the LLM. You can add text instructions and resources, such as retrievers, Flows, customer input, or other resources. A prompt template inherits characteristics from a prompt template type, which contains hidden pre- and post-processing instructions that affect how the resolved prompt is used at runtime. In this example, you create a new prompt template based on the Answer Questions with Knowledge prompt template.

1. From Setup, in the Quick Find box, enter Prompt Builder, and then select Prompt Builder.
2. Click New Prompt Template.
3. For Prompt Template Type, select Knowledge Answers.
4. Name your prompt template (for example, Answer Questions With Case Data) and click Next.
5. Enter instructions into the prompt. Copy and paste from the following example, which comes from the Answer Questions With Knowledge prompt template.

   
               ### INSTRUCTIONS
               1. Analyze the query: Carefully read and understand the user's question or issue from the QUESTION section.
               2. Search CASE: Review the provided company CASE to find relevant information.
               3. Evaluate information: Determine if the available information in the CASE section is sufficient to answer the QUESTION.
               4. Formulate response: To generate a reply <generated_response> to the user, you must follow these rules
               a. Find the article-chunk(s) that is most relevant to answer the user query and VERBATIM extract the ID of the article to set <source_id> field in the response JSON. If you are unable to find the relevant article, set <source_id> to NONE.
               b. Use the relevant article-chunk to generate the response that exactly answers the user's question and set the <generated response> field.
               c. If the user request cannot be answered by the case information provided, set the <source_id> to NONE and <generated_response> to "Sorry, I can't find an answer based on the available articles."
               5. Refine and deliver: Ensure your response is polite, professional, and concise.
               6. Review response: Make sure that have followed all of the above instructions, respond in the desired output format and you only strictly stick to the provided CASE only to formulate your answer.
               
               
               Sample_Case_Retriever_1Cx_cS3a95815db
               
               
               
               ###
               CASE:
               [Insert Retriever Here]
               
               
               
               
               ###
               QUESTION:
               {!$Input:Query}
               
               
               ###
               Response:

6. Delete [Insert Retriever Here] and leave your cursor at that spot.
7. Click Insert Resource, select Einstein Search > Case, and add your retriever.
8. In the retriever settings panel:
   1. For Search Text, select Free Text, then Query.
   2. For Output Fields, select Chunk.
   3. For this example, set Number of Results to 5.
9. In the prompt template, delete the InputQuery under QUESTION and leave your cursor in that spot.
10. In the settings panel on the left, under Search Parameters, click Search Text, select Free Text, Query, and copy the query input (Input:Query). Under QUESTION:, paste Input:Query.

    Input:Query

    
11. Save your work and then click Activate.

Step 10: Test Your Prompt Template​

Test the prompt template you just created.

1. Click Preview.
2. In Preview Settings, expand Inputs if needed.
3. Enter a question into the Query, such as What are some common issues with the toaster oven product?
4. For Retriever ID, add your retriever ID. (In AI Models, select Retrievers, select your retriever and, on the retriever details page, copy your retrievers’s API name, then paste it in the Retriever ID field).
5. Click Preview to see the LLM’s response.

   Preview Shows the Prompt Request and the LLM Response

   

   • Resolved Prompt shows the resolved prompt that was sent to the LLM. The resolved prompt contains the chunks of data that your retriever fetched. That’s RAG in action!
   • Generated Response shows the prompt template’s output that you can surface to customers.

Note If you get an error, try again with the same question or a slightly different one. You can also try refreshing your browser window.

Step 11: Build an Agent that Answers Questions Using Case Knowledge

Finally, create an agent that uses your prompt template to answer questions using your case data.

Tip Developers: You don’t need to consume your prompt template from an action. Developers can consume it from Apex or a Flow instead.

1. Create an Agent Action​
   1. From Setup, in the Quick Find box, enter Agentforce Assets, and then select Agentforce Assets.
   2. Click the Actions tab.
   3. Click New Agent Action.
   4. For the Reference Action Type, select Prompt Template.
   5. For the Reference Action, select your prompt template. If you don’t see your prompt template, make sure it’s active.
   6. (Optional) Name your action something like Answer Questions With Case Data, then click Next.
   7. Add Agent Action Instructions. For example: Answer questions that customers have about product issues. Use the provided customer cases, where the description and subject tell you what is going wrong, or how the customer is struggling. For example: “What are some common issues with the toaster oven?”
   8. For Inputs, fill in the required fields.
      Examples:

      Loading Text: Getting an answer…

      Inputs instructions: Use the retriever that is hard coded in the prompt template.

      Query Instructions Required. A string created by generative AI to be used in the Case search.

      Collect data from user: (make sure it's selected under Query Instructions)

      Citation Mode Instructions: Do not generate citations.

      Accept all other defaults.
   9. For Outputs, for the Prompt Response, enable Show in Conversation.
   10. Accept all other defaults and click Finish.
   11. Select the created agent action to view its details.

       Agent Action Details

       
2. Create an Agent
   1. From Setup, in the Quick Find box, enter Agentforce Agents, and then select Agentforce Agents.
   2. Click New Agent, select Agentforce Service Agent, and click Next.
   3. Deselect everything except General FAQ, then click Next.
   4. For company, add A company that sells toaster ovens, accept the other defaults, and click Next.
   5. Skip the data library and click Create.
      

      Note If you see an error, wait a minute or two and check whether the agent was created. If not, try refreshing the browser window or repeating the instructions in this step.
3. Add a New Subagent to Your Agent​
   Create a new subagent to consume the action from the previous step.
   1. In your agent, select New -> + New Subagent.
   2. For what you want this subagent to do, enter Answer questions with Case information, and click Next.
   3. For Classification Description, replace the default value with something like : This subagent addresses user inquiries related to toaster oven cases and provides relevant details.
   4. For Scope, accept the default instructions.
   5. For Instructions, replace the first instruction with: Gather as much information as possible from the user by answering clarifying questions.
   6. Delete all the other default instructions and click Next.
   7. Select the action you created earlier (Answer Questions with Case Data), then click Finish.
   8. Click Activate to activate your agent.
   9. If you see the Agent Readiness Checklist, click Ignore & Activate.
      The best practice is to review and resolve any issues. However, that's not needed for our example.
   10. Test your agent. In the Conversation Preview window, ask a question such as What are some common problems with toaster ovens?

       Test Your Agent

       

       Note that the answer is not a generic answer about all toaster ovens. Instead, it reflects the information you entered as case subjects and descriptions. It’s easy to see how the LLM response is grounded in the very data source you populated.

Possible Next Steps