Building an Event Assistant Agent in 5 Days with Agentforce and Data Cloud RAG

salesforce · Tech · 2024 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Event attendee assistant

Salesforce needed to migrate a homegrown Dreamforce event agent onto Agentforce and Data Cloud quickly enough to ship before the event, while improving grounded FAQ, schedule, and session-recommendation behavior.

Constraints:

The team had only five days to reach a minimum viable product.
Session search needed both semantic understanding and lexical precision for names.
The agent had to connect to the Salesforce Events mobile app and user schedules.
The team needed rapid debugging and feedback while operating under an event deadline.

Architecture

Salesforce migrated Ask Astro from a homegrown OpenAI/function-calling/vector database setup to Agentforce Service Agent plus Data Cloud RAG. Event session data was ingested from CSV into Data Cloud, indexed with unstructured semantic search and hybrid search, exposed through topics and invocable actions, connected to the mobile app through Bot API, and monitored with Einstein feedback.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  attendee[Attendee] --> salesforce-events-app[Salesforce Events App]
  salesforce-events-app --> ask-astro-agent[Ask Astro Agent]
  ask-astro-agent --> agentforce-service-agent[Agentforce Service Agent]
  ask-astro-agent --> data-cloud-rag[Data Cloud RAG]
  data-cloud-rag --> hybrid-search-index[Hybrid Search Index]
  ask-astro-agent --> invocable-actions[Invocable Actions]
  ask-astro-agent --> einstein-feedback[Einstein Feedback]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Attendee`	Dreamforce app user asking event questions and managing schedule.	`Salesforce Events App`
`Salesforce Events App`	Mobile app surface for Ask Astro.	`Ask Astro Agent`
`Ask Astro Agent`	Agentforce event assistant for FAQs, schedules, and session recommendations.	`Agentforce Service Agent`, `Data Cloud RAG`, `Invocable Actions`
`Agentforce Service Agent`	Configured agent runtime with topics and actions.	`Ask Astro Agent`
`Data Cloud RAG`	RAG layer over event FAQ and session data.	`Hybrid Search Index`
`Hybrid Search Index`	Combines semantic and keyword search for session retrieval.	`Data Cloud RAG`
`Invocable Actions`	Actions for schedule updates and app integrations.	`Ask Astro Agent`
`Einstein Feedback`	Feedback collection for monitoring and improvement.	`Ask Astro Agent`

Data flow

Dreamforce session and FAQ data is ingested into Data Cloud, indexed for semantic and hybrid search, and exposed through Agentforce topics and actions. The mobile app sends attendee questions to Ask Astro, which retrieves grounded FAQ/session context, manages schedules through invocable actions, recommends sessions, and records feedback through Einstein feedback.

“Next, I created a vector index using Data Cloud’s unstructured data semantic search, and headed out for a meal break. When I returned, the vector index was already waiting for me.”article, stated

“Ask Astro Agent helped attendees find useful information grounded in our FAQs, managed user schedules, and recommended sessions”article, stated

Why this architecture

Salesforce chose Agentforce and Data Cloud RAG to compress delivery time and avoid rebuilding agent orchestration, retrieval, app integration, and feedback plumbing by hand. Hybrid search was added because vector search alone failed on precise speaker-name matching.

“The Marketing Event Tech team has been working on migrating a homegrown AI agent to the Agentforce stack since Aug. 26, and we could use your help to add the retrieval augmented generation [RAG] feature to it.””article, stated

“Optimizing name searches was way harder than I thought, though — partly because the speaker name was only a small part of the session information that was being vectorized, and there were a lot of speakers named Adam or Adams. Vector search alone wasn’t going to cut it.”article, stated

Models used: Agentforce Service Agent · OpenAI function-calling · LLM-assisted Java generation

Scale

The MVP window was five days before Dreamforce.
“ ## Minimum viable product: 5 days to Dreamforce”article, stated
The Agentforce configuration included three topics and 12 actions.
“After the vector database calls in Apex classes were implemented and Apex actions were added, an Ask Astro Agent with three topics and 12 actions quickly showed up in the system, and began responding well to test utterances.”article, stated
The debugging loop could be as short as 30 seconds.
“With LLM-powered agents and our Agent Builder, the debugging cycle was as short as 30 seconds.”article, stated
The agent shipped to colleagues and Dreamforce attendees the next day after green-lighting.
“ship the Data Cloud version of Ask Astro Agent to our colleagues, and to Dreamforce attendees on the following day”article, stated

Alternatives considered (roads not taken)

Keep the homegrown OpenAI function-calling and Heroku vector database stack.

Rejected because: The team migrated the agent logic to Agentforce and Data Cloud to meet the deadline.

“We migrated a homegrown AI application from an open-source stack to a Salesforce stack in a matter of days, using all the latest technology that Agentforce and Data Cloud have to offer: * Mulesoft Anypoint Studio * Data Cloud Ingestion API * DLO => DMO mapping”article, stated

Use vector search alone for session retrieval.

Rejected because: Speaker-name accuracy needed lexical precision as well as semantics.

Delay monitoring until after launch.

Rejected because: Einstein feedback was turned on at Dreamforce to observe interactions and improve FAQ content.

“By turning on Einstein feedback, we were able to track how attendees were interacting with Ask Astro Agent, and understand where to improve our FAQ.”article, stated

Patterns

`rag-retrieval` (event knowledge)

Ask Astro grounds FAQ and session responses in Data Cloud RAG.

“add the retrieval augmented generation”article, stated

`hybrid-search` (retrieval)

Hybrid search combines semantic and keyword search to improve speaker-name queries.

“Hybrid search offers the best of both worlds”article, stated

`tool-use-orchestration` (agent actions)

Agentforce topics and invocable actions connect the agent to schedule updates and app behavior.

“three topics and 12 actions”article, stated

`observability-tracing` (feedback)

Einstein feedback records attendee interactions for improvement.

“track how attendees were interacting with Ask Astro Agent, and understand where to improve our FAQ”article, stated

Tradeoffs

Platform migration

Chose Agentforce and Data Cloud RAG over Homegrown OpenAI/function-calling and Heroku vector database stack.

Rationale: The platform stack compressed configuration, retrieval, actions, and feedback into a shippable five-day path.

“migrating a homegrown AI agent to the Agentforce stack since Aug. 26”article, stated

Retrieval strategy

Chose Hybrid search over Vector search alone.

Rationale: Semantic retrieval needed lexical support for precise speaker names.

“Keyword search, on the other hand, excels at lexical similarity but not semantic similarity.”article, stated

Operational scope

Chose MVP in five days over A slower custom rebuild.

Rationale: The event deadline forced an MVP that reproduced current status on the Data Cloud stack.

“see if we could create a minimum viable product (MVP) to replicate the current status of Ask Astro”article, stated

Gotchas (what bit them)

Vector similarity can fail exact-name search.

Root cause: Speaker names were a small part of vectorized session records and similar names collided.

Resolution: Add hybrid search to combine semantic retrieval with keyword precision.

“speaker name was only a small part of the session information that was being vectorized”article, stated

Deadline-driven agent work needs very short debug loops.

Root cause: The team had to tune and ship before Dreamforce with multiple workstreams active.

Resolution: Use Agent Builder and platform feedback cycles that made debugging as short as 30 seconds.

“Testing, FAQ content updates, and small prompt adjustments continued throughout the day across all workstreams.”article, stated

Evaluation

Salesforce used teammate Slack feedback, rapid Agent Builder debugging, production green-lighting, and Einstein feedback after launch to validate and improve Ask Astro.

Metrics: Time to MVP · Debug cycle time · Speaker-name retrieval accuracy · Attendee interaction feedback · FAQ improvement signals

“saw feedback from my teammates in Slack about how the bot was 80% there already”article, stated

Outcomes

Ask Astro went live in the Salesforce Events app on the first day of Dreamforce.
The final architecture used Agentforce Service Agent, Data Cloud RAG, hybrid search, invocable actions, Bot API, and Einstein feedback.
The team improved configurability, debugging, and productivity compared with the open-source stack.

The source reports a five-day MVP window, three topics, 12 actions, a 30-second debugging cycle, and launch to Dreamforce attendees.