You will learn how Salesforce assesses knowledge content before indexing, how it tests Data 360 retrieval changes against a golden data set, and why a human-reviewed failure rate matters more than raw metrics.
Hypothesize how the change affects answers
Build the recipe in Data 360
Evaluate with quantitative and qualitative tests
Reject or deploy, then keep monitoring
The loop used to test each change to the agent's Data 360 configuration
The Salesforce Help Agent is grounded in Salesforce’s own product knowledge, prepared and organized in Data 360 and used by Agentforce through retrieval augmented generation. The authors argue that an agent’s answers can be no better than the content it retrieves, so content quality has to be checked before any model or configuration work starts.
The first lesson is a Fit, Gap, Clash review of every article: whether it answers a real question, whether it leaves a customer issue uncovered, and whether it contradicts other content. Articles are scored using health, failure, resolution and traffic signals, paired with top case drivers. More content does not mean better coverage, and the team chose a separate agent for the Informatica help portal rather than adding that content to the existing Help Agent, which now resolves 92% of questions on that site without a case being filed. Content authors and an AI and Data team share the work, with Agent Managers connecting them.
The second and third lessons describe building by hypothesis: each change to a Data 360 recipe is tested, and only changes that improve answers ship. A product-name metadata hypothesis showed flat results and was rejected, leading to a new recipe that pairs metadata with prompt instructions. Evaluation uses a golden data set of about 200 utterances, each validated by a person, with acceptance criteria covering correct, complete and relevant answers.
The final lesson combines deterministic retrieval metrics, semantic quality checks and business outcomes with human and LLM review. When section-aware chunking was tested against fixed-size chunking on 195 query-document pairs, metrics flagged 8 failures, but review showed the real failure rate was closer to 1%, and that figure was used to approve production.
By Vaibhav Raizada and Kumar Kasimala · Salesforce EngineeringIt is useful if you are designing any AI feature where model output must remain safe and schema-valid.
This explains how Salesforce keeps LLM-assisted prompt creation deterministic: graph-controlled routing, server-side record identity mapping, and sentinel-tagged structured output.
Why read this
It is useful if you are designing any AI feature where model output must remain safe and schema-valid.
Salesforce engineers describe building Assistive Authoring for Prompt Builder, which turns a plain-English requirement into a working Prompt Template in under a minute instead of the roughly 25 minutes an experienced admin needs today. The core risk they targeted was not obviously bad output but plausible-looking templates that break at runtime because the model invented a field, corrupted a record ID, or returned malformed metadata.
The solution uses Agent Graph to split the work into a router node and three specialized sub-agents for generating, refining, and clarifying requests. State variables carry client-supplied and runtime context, bound inputs pass only what each tool needs, and four backend tools run in a fixed sequence with a beforeReasoning hook that fetches grounding before the model loop starts, cutting latency by removing a model decision from the critical path.
Deterministic boundaries protect the three failure points. The model can signal intent through a state variable, but the graph evaluates guards before changing nodes. The model never touches record IDs; it writes stable developer names and a server-side transpose layer maps them back to real records and IDs. Template metadata is wrapped in sentinel tags, the client extracts and parses only that payload, and parse failures return errors rather than guesses.
Grounding is handled by a discovery service that normalizes SObject fields, flows, Apex-backed providers, and search retrievers into one consistent shape, runs in the current user’s context with field-level readability filtering, and reduces the result to an allowlist the model must select from. Safety is enforced by making output a draft only: the agent has no tool to save, publish, or execute, content becomes typed editor blocks with escaped interpolation, and the admin accepts or rejects a diff before anything is saved.