Grounding Agents with Data 360 and Trusted Public Sector Data
Grounding, precisely
Grounding means retrieving relevant, permission-respecting data at conversation time and placing it in the prompt. The model is not retrained; nothing is learned permanently. That single distinction answers a large share of exam questions.
The Data 360 path
Data is ingested through connectors and data streams (batch, streaming, or zero-copy), mapped to the data model, harmonised, and often resolved through identity resolution into a unified profile. For unstructured material - policy PDFs, procedure manuals, statute summaries - you index and chunk the content, then retrieve the best-matching passages semantically. Structured material (case history, 311 service requests, permits, licence status, eligibility records) is usually retrieved through an action that queries the object or a calculated insight.
Trusted public sector data
"Trusted" is about provenance and permissions, not volume. Prefer curated Knowledge articles, governed Data 360 objects, and system-of-record queries over scraped or unvetted documents. Retrieval runs in the context of the user or agent runtime, so sharing rules, data spaces, and field-level security still apply.
Traps
- Retrieval-augmented generation is not fine-tuning and does not make the agent "remember" between sessions.
- Grounding does not bypass security - if a caseworker cannot see a record in the UI, the agent should not surface it either.
- Stale or unindexed content produces confident-sounding wrong answers; refresh cadence is a design requirement.
- Identity resolution errors look like model errors: right answer for one constituent, wrong for another.
Rule of thumb
If the answer is wrong for everyone, suspect the source data. If it is right for one user and wrong for another, suspect identity resolution or permissions - not the model.