Data Foundations and Grounding for Agents
Grounding defined
Grounding is the practice of injecting relevant, permission-aware context into the model's context window at runtime. In Agentforce, grounding comes from CRM records, Knowledge articles, Prompt Builder templates, and Data Cloud retrievers over structured and unstructured data. Data Cloud ingests streams into data lake objects, maps them to data model objects, resolves identity, and builds search or vector indexes for retrieval. Retrievers chunk, embed, and rank content so the agent can use authoritative context.
Concepts to know
- RAG vs. fine-tuning: Agentforce grounding is primarily retrieval, not model training.
- Dynamic grounding: context is fetched per request, not baked into the prompt.
- Permission-aware retrieval: the agent user's access filters what can be retrieved.
- Freshness and chunking directly affect answer quality.
Traps
Candidates assume uploading a PDF is enough; it must be ingested, indexed, and attached to a retriever or topic. Another trap is thinking grounding bypasses sharing rules. More context is not better: irrelevant chunks dilute answers. Do not confuse Data Cloud data streams with retrievers.
Worked example
A product-support agent needs warranty-policy answers. Load PDFs into Data Cloud, chunk and index them, create a retriever, attach it to the Warranty topic, and test with paraphrased questions. Rule of thumb: ground before you prompt-retrieve the smallest authoritative chunk, then generate.