Agentforce RAG (Retrieval-Augmented Generation): Architecture and Implementation Best Practices - Oct 21, 2025

Key Takeaways To Expect At The End Of This Session:





Design smarter embeddings — why custom chunking strategies can make or break accuracy when handling complex technical documents.



Optimize retrieval performance — how hybrid search (keyword + semantic) outperforms pure semantic search when paired with structured metadata.



Craft effective prompts — understand how prompt design and chunk size consistency impact the quality of your RAG responses.



Establish robust feedback loops — learn how to evaluate, tune, and prevent RAG quality from degrading over time.



Balance tooling and customization — when to rely on Agentforce Data Library accelerators, and when a custom prompt template delivers better results.

1. Trailblazer Community
Template2024
Last Edited: 4/23/2024

2. Varsha Tiwari
Senior Principal,
Client Delivery Enterprise
Nicolas Lebel
Senior Principal,
Client Delivery Enterprise

3. AgentforceRAG
Architecture and ImplementationBest
Practices
Nicolas Lebel
2025-10-21, Salesforce Architecture User Group, Montreal

4. RAG definition
The process of optimizingthe output of a large languagemodel, soitreferencesan authoritative
knowledgebase outsideof itstraining data sources beforegeneratinga response. Large Language
Models(LLMs) are trainedon vastvolumes of data and use billions of parametersto generate
original output for taskslike answeringquestions, translatinglanguages, and completing
sentences. RAG extendsthe alreadypowerfulcapabilitiesof LLMsto specificdomainsor an
organization'sinternalknowledgebase, all withoutthe needto retrainthe model.
Source: https://aws.amazon.com/what-is/retrieval-augmented-generation/

5. Manydifferentcompanycontextswhenitcomesto documentation
○Accumulatedcontent in multiple sources
○Mature Salesforce Knowledgeusers
○Differentdocuments types(PDFs, Word docs, HTML) or Salesforce records, or a combination?
○Document content: FAQ, manuals, Conversations, Case resolutions, etc…
○Type of audience for yourcontent

6. The AgentforceAgent Flow

7. RAG architecture: All about Offline Preparation

8. Initial content optimizations
User manualsand theirlimits
●Typicaluser manual, the semanticmatch betweenquestions and content islessclear.
●If youcan authoryourcontent withQ&A, itwillgetvectorizedand leveragedby the LLM for Context.
●Lots of structureddata insideuser manuals(thinkof tables). Need to givecontextof the information.
○Withoutthe use of TabularEmbeddingModels, embeddingtabularinformation isa nightmare.
○Right now, withAgentforceand Data 360 (DataCloud), the recommendationisto extracttables intoJSON or HTML files.
●Content governanceneedsto beimplemented
○LeverageCMS, training, productteams to makesure content isup to date
○A properlymanagedSalesforce KnowledgeBase ismore manageablethan5000 PDFssittingin a folder.
Content relevance and reliability
Load Chunk Vectorize Index Retrieve

9. Chunkingstrategies
●Section-AwareChunking
●Semantic-basedPassage Extraction
●Conversation-basedChunking
●PrependField Chunking
Load Chunk Vectorize Index Retrieve

10. Section awarechunking
Load Chunk Vectorize Index Retrieve
●Default mode for embeddedPDF and HTML files whenusingData 360 advancedbuilder
●HTML tags(H1-H6) strippedoff by default
●Max token: whendocuments have short paragraphsor listitems
●Overlaptokens: Contextwindowaroundthe chunk

11. Semantic-basedPassage Extraction
Load Chunk Vectorize Index Retrieve

12. Conversation basedchunking
Load Chunk Vectorize Index Retrieve

13. PrependField Chunking
Use additionalfieldsor metadatato providecontextfor a chunk
512 tokenlimit
Ex: Description fieldon Knowledgearticle
How to use prepending
1.Access data preparationsettings:Navigateto the settings for your
unstructureddata in Salesforce Data Cloud.
2.Select a chunkingstrategy:Choosethe appropriatechunkingstrategybasedon
yourdata.
3.Enable fieldprepending:Select the "Prependfieldsto eachchunk" option.
4.Addfields:Select the fieldsyouwantto prepend.For example, youcouldselect
"ProductName" and "ProductSKU" to addto eachchunkof productdescription
data.
5.Proceedwithvectorization:Continue the process by selectingan embedding
model for vectorizationand indexing.
Load Chunk Vectorize Index Retrieve

14. Enrichedindexingor chunking
●Especiallyusefulin UDMOswherefieldprependingisnot possible
●Plain, Question, Metadatachunktypes
●Chunk enrichmentincreasescostand latency
●Open AI ADA 002 needto beused
oNot supportedon MultilingualE5-Large or E5-Large V2
Load Chunk Vectorize Index Retrieve
ContextIndexing(justreleased)
●Be able to analyse how the chunkinggetsexecutedfor one specificdocument

15. Differentoptions for RAG in Agentforce
●AgentforceService Agent / Agentforcefor Employees
○Answerquestions withKnowledgeUsingADL
■Knowledgearticles
■Uploadedfiles
○Use custom prompt templatewithADL
○Use custom prompt templatewithcustom connector
Load Chunk Vectorize Index Retrieve

16. The limitsof AgentforceData Library (ADL)
●SearchIndex
○512 tokensper chunk
○E5 Large Multilingualembeddingmodel
○Hybridsearchmode
○No enrichedchunkingfor now
●Retriever (1 per ADL withfilteron Source Id)
○Returns10 resultsby default
○Advanced retrievalmode isOFF
Load Chunk Vectorize Index Retrieve

17. The resurganceof KnowledgeArticles for RAG
●Whenchunkingand vectorizingSalesforce knowledgearticles, the searchindex isbuilt
againsta structuredDMO. Takeadvantageof thisstructure by spreadinglong-textcontent
acrossmultiple fields, suchas Question, Description, Resolution, and Exceptions. Annotate
the knowledgearticle withmetadatafor filteringand prepending
Load Chunk Vectorize Index Retrieve

18. Ensemble Retriever
●Combine multiple retrievers to dynamicallyreranksearchresultsfor a sameprompt template.
●Onlyone retriever isusedto groundthe prompt template.
●The mostrelevant queryresultsare surfacedat the top of the ranking
●No irrelevantresultsare addedto the prompt.
●The prompt consumes fewerEinstein Requests, whichreduceslatencyand cost.
Load Chunk Vectorize Index RetrieveLoad Chunk Vectorize Index Retrieve

19. Multi-languageuse cases
●You couldembedcontent in manydifferentsupportedlanguagesand have the usersprompt in otherlanguages
●Need to thinkof
○Trust layer: PII information configuration
○Supportedlanguagesin the Embeddingmodel
○LLM model supportedlanguages(bothfor input and output).
○AgentforceLanguages
○Reasoningengine Languages
Load Chunk Vectorize Index Retrieve

20. Key links
Both authored by Reiniervan Leuken,Senior Directorof Product
Management-Agentforce
https://www.salesforce.com/agentforce/agentforce-and-rag/
https://www.salesforce.com/plus/experience/tdx_london_2025/series
/tdx_2025_london_highlights/episode/episode-s1e5

21. Q&A

22. Quick Survey