Choosing an agent
You select an agent in two places:- Home page — The agent dropdown next to the message input.
- Sessions page — The agent selector when creating a new session, or the agent switcher at the top of an active session. Press
⌘⇧G(orCtrl+Shift+G) to open it from anywhere.
Available agents
The agents available to you depend on your account’s data connections and configuration. Common agents include:UNFI (Natural and Conventional)
Analyzes UNFI distribution data split into two channel-specific agents — Natural and Conventional — each backed by its own semantic model. Supports:- Assortment gap and whitespace analysis
- Inventory and weeks-of-supply assessment with stock-out and overstock risk
- Daily and weekly transaction data
- Store and DC-level detail
- Retailer-level access controls and supplier-intersection filtering
Built-in capabilities
Every agent shares a common set of capabilities. The specific tools available depend on the agent and your account configuration.- Query data — Run queries against semantic models (Malloy or XMLA) and return structured results. UPC/EAN filter values are automatically normalized to match the target column’s format, so you can paste a barcode in any format and get correct results.
- Analyze and transform data — Run calculations, prepare datasets, and generate files as part of an answer.
- Create visualizations — Produce D3 charts via a specialized charting capability with access to the full charting library, styled for your theme.
- Generate presentations — Produce formatted
.pptxdecks from your data. See Presentations. - Generate files — Create CSVs, markdown reports, and PDFs.
- Edit existing files — Make targeted search-and-replace edits, search inside a file before editing, and diff two files. Changes appear as inline diff cards in chat.
- Schedule work — Turn any conversation into a recurring or one-off scheduled run. Ask an agent to “run this every Monday at 8 AM and email me the results.” See Schedules.
- Search documentation — Look up answers about the UNFI Insights platform, data definitions, and common workflows from the built-in knowledge base.
- Send emails — Deliver results via email with formatted content and attachments.
- Search the web — Look up external information using Google search.
- Browse web pages — Open a page, extract tables and content, and follow links using an autonomous browser agent.
Advanced analytics skills
All standard agents include a set of advanced analytics skills that guide rigorous multi-step workflows:- Price elasticity screening — Observational demand-response playbook with store/product/week grain rules and Python diagnostics. Framed as associational historical screening, not causal price recommendations.
- Diff-in-difference (DiD) — Store-level intervention impact with 2×2 DiD / TWFE and cluster-robust standard errors. Detects staggered rollouts and stops with guidance instead of silently switching estimators.
- Store clustering — Lighter multi-phase playbook that requires a cluster-profile table in the report, defers heavy artifacts, and includes a fit-check exit path when the user’s ask is not clustering-shaped.
- Assortment gaps — Store-level void scoring for assortment whitespace within your sales footprint: detects voids where a product sells in some of your stores but not others, combines multi-level affinity indexes, and returns dual rankings (best-fit stores by estimated velocity, plus a de-risked score that favors proven distribution). Ask things like “where should we expand this product” or “find assortment gaps for this category.” Scoped to stores where you have sales history, so it surfaces gaps within your existing footprint rather than the full universe of category-carrying stores. Framed as a scoring heuristic, not causal, so cannibalization, shelf space, seasonality, pricing, and supply chain are not modeled.