Chat¶
The hub of Adverity Atlas. Ask questions about your data, create automations, generate artifacts, everything starts here. No SQL required.
Prerequisites¶
At least one data source connected
Glossary reviewed (recommended, improves answer accuracy)
An AI provider configured (the platform-managed default works out of the box)
The Analyst role or above. A Viewer can read conversations but not start one, see Members and roles
How it works¶
You ask a question → the glossary resolves terms and metrics → SQL is generated and executed → results come back as charts and tables.
Ask a question in plain language
Resolve terms and metrics from the glossary
Generate and run SQL against your warehouse
Answer as charts and tables
Conversations are multi-turn. Each follow-up inherits context, “break that down by country” or “now just for last week” works naturally.
Beyond asking questions, you can get into creating mode: build automations, generate artifacts, edit reports, all from the same conversation.
Note
Adverity Atlas executes queries with read-only access. It does not write to your tables, modify data, or create tables.
How Chat fits the bigger picture¶
Chat sits at the center of Adverity Atlas. What comes before and after a conversation is as important as the conversation itself:
Before: Connect a data source so there is data to query, then review the glossary and add knowledge so answers use your organization’s definitions
During: Ask a question, refine it with follow-ups, and narrow scope with @mentions
After: Save an answer as an artifact to share it, or turn a conversation into a reusable automation that runs on a schedule
Asking questions¶
Type a question in plain language and send it. Adverity Atlas returns a visual answer, a chart or a table, with the result traced back to the query that produced it. The clearer the question, the more accurate the answer.
Questions tend to fall into a few types:
Performance: “What was our ROAS by channel last month?”
Comparison: “Compare cost per acquisition across Meta and Google for Q2 2026”
Anomaly: “Which campaigns had a sudden spend increase last week?”
Exploration: “Which audiences drove the most conversions in June?”
Follow-ups build on the previous answer, so you can start broad and drill in: ask for ROAS by channel, then “just paid social”, then “compare that to last quarter”.
Mentioning data¶
Type @ to reference something specific in your question. This narrows the scope and helps
Adverity Atlas pick the right data. Two things can be mentioned:
Mention |
What it does |
|---|---|
A data asset |
Points the question at a specific table or dataset from a connected source |
A file |
Includes an uploaded file in the analysis |
Tip
Type @ and start typing; Adverity Atlas suggests matching data assets and files as you go.
Attaching something to a conversation¶
An @mention points at data that is already in the workspace. When what you want to discuss is on your clipboard or on your desktop, attach it to the conversation instead. There are two ways in:
Attach by |
What happens |
|---|---|
Dropping files on the conversation |
A “Drop files to attach to this chat” overlay appears; releasing stages the files |
Pasting |
A long block of text becomes a text attachment instead of filling the composer, and a pasted image is staged as a picture |
Anything you attach is staged first, not uploaded immediately. Each staged item shows as a pill above the composer with its filename and size, marked “pending, will upload on send”. Click the pill to preview it, or the ✕ to drop it again. The upload happens when you send the message, so a mistaken paste costs nothing.
Once sent, the files move out of the composer and into the message they were sent with. An image shows as a thumbnail you can click to preview, and any other file shows as a chip that downloads it. Hover a file for two more controls:
Control |
What it does |
|---|---|
Download |
Downloads the file |
Remove from conversation |
Takes the file out of the conversation’s context. It stays in the message, dimmed, as a record of what was sent. Not offered on a conversation shared with you |
Note
Short pastes behave normally and go straight into the composer. It is only a long block, over roughly five thousand characters, that becomes an attachment, which is what stops a pasted report from swamping the message box.
Stopping a response¶
While Adverity Atlas is working, the send button becomes Stop. Click it to end the current response rather than waiting it out, then rephrase and ask again. This is the one to reach for when you realise mid-answer that the question was wrong.
Answering a form¶
When Adverity Atlas needs several decisions from you before it can go on, for example the settings for a model fit or a budget plan, it posts a form in the conversation instead of asking one question at a time. Each field is already filled in where Adverity Atlas could infer a sensible value from your data, so most of the time you are confirming rather than typing.
Field type |
What you do |
|---|---|
Text |
Type a value, in a single line or a larger box |
Number |
Enter a number |
Date |
Pick a date |
Yes/no |
Switch an option on or off |
Choice |
Pick one option, or several where the field allows it |
Range |
Set a value on a slider, or a Low and a High bound |
Workspace or data source |
Pick from the workspaces you can access or the data sources in this workspace. Some fields let you add more than one with + Add another |
Review each field and change anything that is wrong.
Click the button at the bottom of the form, Submit unless the form names it differently. Adverity Atlas checks each answer against the field, for example that a required field is filled in, and marks the field when an answer does not fit.
Adverity Atlas posts your answers as your reply and carries on with them.
An answered form stays in the conversation, collapsed to the values you gave.
Approving what Adverity Atlas saves¶
When you ask Adverity Atlas to remember something (“our ROAS target is 3x, remember that”) or to correct an item in your knowledge base, it does not save anything on its own. It shows a card with exactly what it would store, the type, name, summary, content, what the item is linked to, and when it is used, and the note “Choose for each item, then submit. Nothing is saved until then.”
Approve: Marks the item to be saved as shown
Reject: Marks the item to be left out. Adverity Atlas does not try again unless you ask
When a reply proposes several items, each gets its own card, so you can approve some and reject others.
Click Approve or Reject on each card. Your choice is recorded straight away and you can change it until you submit. A Submit button below the last card counts how many items are decided, for example “1 of 2 decided”.
Click Submit once every card has a decision. The button stays unavailable until then.
Adverity Atlas posts your answer as your reply, for example “Approved 1, Rejected 1”, saves the approved items, and carries on. Each card is marked Approved or Rejected.
If you send a new message instead of submitting, the cards you already decided keep your choice, and any card without a decision saves nothing and is marked Not saved. A new item saved this way arrives in the knowledge base as Detected, ready for review like any other proposed item.
Adverity Atlas saves knowledge only when you ask it to, never to keep a conclusion of its own. Saving and changing knowledge needs the Editor role. Without it, no card appears and nothing is saved.
Confirming actions after reading outside content¶
Content from outside your warehouse, such as a web page, an uploaded file, a past conversation, or a result from an MCP server, can contain instructions written to steer an AI. So once a conversation has read content like that, Adverity Atlas asks before it takes a consequential action: one that sends data out of Adverity Atlas, has an effect that outlasts the turn, or makes a destructive or privileged change. Outbound MCP tools that their server does not mark as read-only count as consequential.
The request appears as a card in the conversation, the same kind used for saving knowledge. It lists the action’s arguments in full and ends with a Why you are asked line naming the sources, for example “This conversation read content from the web.”
Read the arguments and check they match what you asked for.
Click Approve to let the action run, or Reject to stop it.
A conversation stays in this mode once it has read outside content, so later consequential actions in the same conversation ask too. A tool that reads outside content and acts on it in the same call always asks. Automations have no one to ask, so they skip a consequential action in this situation instead of running it.
Tracing an answer¶
Every answer shows its working. Each step the agent took appears as a pill beneath the answer, labelled with the step and how long it took in seconds. Click a pill to open the Data Provenance panel for that step.
What the panel shows depends on the kind of step:
Step |
What the panel shows |
|---|---|
A warehouse query |
The SQL Query that ran and, under Ran against, the data source it ran on |
A file read |
The sheet that was read, under Ran against |
An analysis step |
The Analysis performed, the Operation, and the DataFrame and Parameters it worked on |
Any step |
Agent Reasoning and Thinking Process, the reasoning behind the step, plus the raw Tool Arguments |
Show technical detail expands the identifiers behind a step, the data source or file id and the Row limit that was applied. Hide technical detail collapses it again.
Rerunning a query¶
A stored result is a snapshot of the moment it ran. To check it against the warehouse as it is now, click Rerun query on a warehouse step. The button appears only for the Admin role. Adverity Atlas re-executes the same SQL and opens a Live query result panel marked Live and “Re-executed just now against the live warehouse”, reporting:
Rows: How many rows came back
Duration: How long the query took
Bytes scanned: How much data the warehouse read
Datasource: Which source answered it
Note
Provenance respects your access. If a data source or file has since been withdrawn from you, the panel says the detail of that step is no longer available to you rather than showing it.
Checking one claim inside an answer¶
An answer is often right in outline and worth pressing on in one place. Select the sentence or figure you want to examine and a small toolbar appears over the selection with three actions:
Action |
What it does |
|---|---|
Verify |
Asks Adverity Atlas to check that claim on its own, without changing the thread |
Copy |
Copies the selected text to your clipboard, confirmed with a “Copied” toast |
Ask follow-up |
Drops the selected text into the message box as a quotation, ready for you to type your question under it |
Verify answers underneath the original message rather than at the bottom of the conversation. The check appears as a Verification: footnote along the message’s left edge, showing the claim being checked and how long the check took in seconds. Click the footnote to expand it for the full response and the step pills behind it, and again to collapse it. Footnotes stay with the message, so a conversation you come back to still shows what was checked.
Several claims in the same message can be verified, each getting its own footnote. Because a verification is a question in its own right, it needs the same Analyst role as asking one.
Tip
Verify is the quicker move when you doubt one number. Reach for Rerun query when you doubt the data underneath it, and a plain follow-up when you want the whole answer redone.
The toolbar appears only on an answer from Adverity Atlas. It does not appear on your own messages, inside a verification footnote, or while a response is still streaming.
Finding past conversations¶
The Chat landing page keeps your history. It has two tabs:
Conversations: Your own threads, with the most recent surfaced under Recent Conversations
Automations: The threads produced by automation runs, listed under Automation conversations
Narrow the list with the search box and filters above it:
Filter |
Options |
|---|---|
Search box |
Reads Search conversations…, or Search automations… on the Automations tab. Matches the title and what was actually said in the thread, so a phrase you remember finds the conversation even when its title does not mention it. A misspelled title still matches |
Category |
All categories, Analysis, Schema, Debug, Reporting, Knowledge, General. Shown on the Conversations tab only |
Time |
All time, Today, This week, This month |
Show archived |
Include archived conversations, which are hidden by default |
Adverity Atlas assigns the category itself from what the conversation is about, so you can come back to “the schema questions from last week” without having tagged anything.
Searching from anywhere¶
You do not have to be on the Chat landing page to find a thread. Search… in the top bar, or
⌘K (Ctrl+K on Windows and Linux), opens a search box prompting Jump to a page, or
search your chats…
Type, and it searches two things at once:
Group |
What it lists |
|---|---|
Workspace |
Data, Glossary, Knowledge, and Chat |
Account |
Your profile, Settings, and Tasks |
Conversations |
Matching threads, each with its title and a line from the thread |
Press Enter to open the first match, or click the one you want. Esc closes the box.
Destinations are matched ahead of conversations, so typing a page name still takes you to the page,
while a phrase you remember from a chat, which matches no page, reaches the conversation.
Working on an artifact beside the conversation¶
When a conversation produces an artifact, a card for it appears in the thread with a Preview badge. Click the card and the artifact opens in a panel next to the messages, so you can see the change and ask for the next one in the same view. Show app in the conversation header reopens the panel after you close it.
The panel’s controls depend on what it is showing:
Control |
What it does |
|---|---|
Dark / Light |
Switches the theme the artifact renders in |
Previous slide / Next slide |
Moves through a presentation, with the slide number between them |
Download |
Saves the artifact. A presentation offers PowerPoint (.pptx) or Web page (.html); other types download a single file |
Full screen |
Opens the artifact on its own, away from the conversation |
Activity log |
For a data app, shows the Activity log of its queries, storage calls, events and errors |
Select an element to edit |
For a data app, lets you click one element of the app and ask for a change to exactly that element. See Data apps |
You can start from the other end too. Edit with AI on an artifact opens a conversation already anchored to it, titled “Editing:” plus the artifact name, so refining something you made last week does not mean finding the original thread.
Creating an automation from a conversation¶
Once a conversation has answers you would want again, turn it into a reusable automation. With the conversation open, click Create Automation. Adverity Atlas reads the conversation and opens a Create Automation panel described as “Extracted from your conversation, review and save”, pre-filled with:
Field |
What it holds |
|---|---|
Name |
A name for the automation. Required |
Description |
What the automation does |
Instructions |
The step-by-step instructions the agent follows, drawn from the conversation. Required, and editable |
Parameters |
“Values for the placeholders in the instructions”, one box per placeholder, pre-filled with the suggested value. Shown only when the instructions use placeholders |
Tools |
The tools the extraction suggests the automation will need |
Trigger |
Scheduled (Cron) or Event-based. Required |
Schedule |
When it should run, shown for a scheduled trigger |
Output |
Post results to chat and Notify me when this automation completes, both ticked to begin with |
Review and adjust anything, then click Save Automation. The automation appears in your automations list and runs on its trigger.
This is how a one-off question becomes ongoing monitoring: the same analysis runs on a schedule and surfaces findings without you asking again.
Note
The button appears once a conversation has messages, and only on your own conversations, not on one shared with you. The extracted automation is a starting point: edit it before saving and tune it later like any other automation.
Changing an automation from a conversation¶
You do not have to open the automations list to adjust one. Ask in the conversation to change an automation’s schedule, a threshold, or the wording of a custom automation’s instructions.
Adverity Atlas answers by showing you the prompt the automation would run after the change, without saving anything. It applies the change only once you say to go ahead. See Creating automations for the same edit through the UI, and for what is editable on a template-based automation versus a custom one.
Defining a glossary term in the conversation¶
When Adverity Atlas does not know a term, you can teach it there and then rather than switching to the Glossary. Describe the term, its meaning, and the columns it comes from:
“Define net_spend as spend after agency fees, with net_cost and spend_net as the source column names”
The concept is created and activated in the same turn, so the next question can use it. Activation also queues the pass that links matching columns, so the concept’s Linked Columns fill in shortly afterwards. If a close match already exists, Adverity Atlas offers it for reuse rather than creating a near-duplicate, and an existing name is routed to an edit instead of a second copy.
Chat can also publish a draft or retire a term that is no longer used, “publish the campaign_type draft” or “retire cost_per_lead”. A draft is invisible to queries until it is published, which includes the drafts Adverity Atlas proposes for itself.
Note
Concepts created this way are scoped to the current workspace, and creating or changing one requires the Editor role. Organization-wide and standard concepts remain an admin action. See Adding concepts.
Statistical analysis¶
Some questions are not answered by another query. “Is that difference real?”, “what drives visits?”, “where do we land at the end of the quarter?” need statistics on top of the rows. Ask for them in the same conversation and Adverity Atlas runs the analysis over the result it already fetched.
Ask for |
What comes back |
|---|---|
A significance test |
A p-value, whether the difference is significant, the effect size, and which test was used. The test is chosen from the shape of the data |
Correlations |
A correlation matrix and the notable pairs, each with its strength and p-value |
Anomalies |
The flagged rows with a severity and the expected range the value fell outside |
A forecast |
A projection with confidence bands, on minute, hourly, daily, weekly, or monthly data, plus which forecasting method produced it |
A regression |
A coefficient table with standard errors, t-statistics, p-values, and confidence intervals, plus R², the F-test p-value, and a Durbin-Watson statistic |
A rolling window |
The series with a moving average, sum, standard deviation, minimum, maximum, or median added as a column |
Regression takes any number of regressors, so a specification such as “visits from TV impacts with up to 10 lags, plus hour-of-day and day-of-week dummies” is a single request. On a time series, ask for robust standard errors so autocorrelation does not make the coefficients look more certain than they are. Fitted values and residuals come back as data, so you can chart them alongside the actuals.
A forecast can cover several metrics at once, one series per value of a column such as campaign (up to 50 series), and take drivers such as spend into account. When you already know the future values of a driver, a planned budget for example, include them and the forecast uses them. Ask for metrics that cannot go below zero, such as clicks or spend, to be kept at zero or above. Where the forecasting model cannot be used, a single metric is forecast with a baseline that repeats the last week, day, or year, and the answer says so.
Each analysis shows in the trace as its own pill, labelled Ran significance test, Checked correlations, Detected anomalies, Forecasted trend, Fitted regression, or Computed rolling window. Click one to see the Operation and the Parameters behind it.
Note
These numbers are computed rather than estimated by the model, so the same data gives the same p-value every time.
Running an analysis needs the same Analyst role as asking a question, since it works on a result a query produced.
Marketing mix modeling¶
For budget-allocation questions that a single query cannot answer, ask Adverity Atlas to fit a Marketing Mix Model on your historical spend and outcome data. The model captures two effects a straight comparison misses: per-channel adstock (spend that keeps working after the week it was spent) and saturation (the point where extra spend in a channel stops paying back).
It returns channel ROAS and contribution share, with a chart, so you can see which channels are carrying the outcome and which are past the point of diminishing returns.
To use it, you need a consistent time series: a date column, one spend column per channel, and the outcome you are modelling. That normally means pulling the data together in the conversation first, then asking for the model.
The data review before the fit¶
Adverity Atlas reviews the data before it fits anything and tells you what it found, check by check: missing or non-finite values, missing or duplicate dates, the cadence and any gaps in it, how much history there is, periods where a channel spent nothing, outliers, columns that never vary, and pairs of channels whose spend moves together. A column with too little variation to be attributed is flagged rather than quietly modelled.
A chart of the outcome and total spend over time renders under the review with the findings drawn on it: outlier periods as red points, gaps as yellow bands, periods with a missing value as dashed lines.
Each finding comes with the fix that suits it, for example filling a channel’s spend with zero for the weeks it was off, inserting the missing periods, or merging two channels that moved together. Missing values are refused until they are fixed. The other warnings only need your decision, and the fit starts once you have made it. Where the data cannot support a fit at all, Adverity Atlas says so and does not start one.
The fit itself runs in the background rather than blocking the conversation. When sampling finishes, the result arrives in the conversation on its own, you do not need to wait on it or ask again. You can also ask about a past fit later.
Ask how well the model fits, or about the saturation and carryover shapes, and Adverity Atlas returns the fit diagnostics: response curves per channel with current spend marked, actual against fitted over the fit window, and carryover per channel.
Optimizing channel budgets¶
Once a fit is complete, ask Adverity Atlas to allocate a budget across its channels for a future window, for example “split 400k across these channels for next quarter”. It uses the fitted saturation and carryover curves to find the per-channel spend that maximizes the expected outcome, and returns the allocation next to current spend, the change per channel, and the expected outcome with its uncertainty interval.
You set the total, the start and end of the planning window, and any limits that matter. By default every channel is kept between half and twice its spend in the fit window, because that is the range the model has actually seen. Widen it for a more aggressive reallocation, hold a channel where it is, or cap one that cannot absorb more. You can also ask it to lean on the channels the model is confident about instead of maximizing the raw expected return, which is the honest choice when several channels have wide uncertainty.
The optimization runs in the background. When it finishes, a Budget plan card appears in the conversation reading “Optimization complete”, and Show plan brings back the allocation with four charts: current against optimized spend over the window, the change per channel, the response curves with both the current and the proposed spend marked, and the spread of the expected outcome for each split. The response curves are the explanation: money moves off channels sitting on the flat part of their curve and onto channels still on the steep part.
The plan is a starting point, not an instruction. The optimizer maximizes the model’s expected response and knows nothing about reach, creative capacity, or contractual commitments you did not give it as limits. Where a channel’s curve was not pinned down by the fit, it is held at its current spend and the plan is reported as directional rather than a recommendation.
Note
Only one budget optimization runs against a fit at a time. Ask for a second while one is in flight and Adverity Atlas points you at the one already running.
Note
Fitting a marketing mix model and optimizing a budget require the Editor role. Anyone with Viewer access can read the results.
Price response modeling¶
When the question is about price rather than media, ask Adverity Atlas to fit a price response model. It estimates, per product, how much volume is lost for every 1% increase in price (own-price elasticity) and how much each promotion tactic lifts sales, such as a retailer flyer, an in-store display, or a temporary price reduction.
Products are modelled together: a product with only a handful of price changes borrows strength from the rest of the category instead of returning a noisy number, and the result flags products whose price barely moved so you know which estimates are measured and which lean on the category.
To use it, you need one row per product per period with the units sold, the net price paid, and 0/1 flags for each promotion. Distribution (the share of stores stocking the product) and media spend can be added as extra drivers.
Before fitting, the agent reviews the data the same way it does for a marketing mix model and tells you what it found, check by check: whether each product has one row per period without gaps, how much history there is, weeks with no sales, unusual weeks, prices that look wrong, whether the price actually moved, and whether each promotion ran often enough and apart from price cuts to be measured. Warnings come with the fix that suits them, and the fit only starts once you have decided how to handle them. It then runs in the background like a marketing mix model and arrives in the conversation as a button when it finishes, with charts of the elasticity intervals per product and the lift per promotion.
The model reports own-price elasticity at the prices seen in your data. It does not say where lost volume goes (to a sibling product or a competitor).
Note
The same roles apply as for marketing mix modeling: Editor to fit, Viewer to read the results.
Researching the web¶
For context that is not in your own data, such as a competitor launch, a platform policy change, or market news, ask Adverity Atlas to look it up on the public web. It can:
Search: Run a web search and read the ranked results
Read a page: Open a specific page you point it at and answer a question about it
Answer: Answer a factual question with its sources
Deep research: Search and read across many pages to produce a written report on a brief
Each step appears as a pill under the answer, for example “Searched the web for” and the query. Open it in the Data Provenance panel to see the sources that were read.
Deep research runs in the background for up to about 10 minutes, so the conversation does not wait for it. The step reads “Deep research running in the background”, with a note that the report will appear in the conversation.
Ask for the research, for example “research how competitors are pricing connected TV this quarter”.
Adverity Atlas starts the research and ends its turn. You can keep working in the meantime.
When the report is ready, it arrives in the conversation as a new message with a summary and the full report as a markdown file, and you get a notification, “Deep research finished:” plus the brief. Reload the page if the message has not appeared.
Ask a follow-up, and Adverity Atlas uses the findings.
The research itself reads only the public web and cannot see your knowledge base. Before starting it, Adverity Atlas looks up the brands, competitors, markets, and terms the task involves in your knowledge base and writes what is relevant into the brief. It keeps that to what the research needs, because the brief can end up in public web search queries.
Only one deep research runs in a conversation at a time. Automations can use web research too; a run waiting for a report shows Waiting for research (see Tracking runs).
Web research needs the Editor role. It uses your AI model’s built-in web search, or a search provider your admin configures under Web research. When neither is available, Adverity Atlas does not offer it.
Checking system status¶
The chat landing page carries a status indicator that reads All Systems Operational, Partial Degradation, or Service Disruption. Hover it for the Subsystem status breakdown: Data Warehouse covers the sources bound to this workspace, LLM Provider covers the model behind Chat. Each line shows its current state and when it was last checked, and the indicator refreshes on its own.
Check it first when answers stop coming back. A degraded warehouse or an unreachable provider is visible here before it looks like a problem with your question.
Asking how Adverity Atlas works¶
Chat answers questions about the product, not only about your data. Ask how to connect a warehouse, what a setting does, or how the glossary tiers resolve, and Adverity Atlas searches this user manual and reads the pages it needs before answering. Those steps appear in the trace as Searched documentation for “…” and Read documentation: …, so you can see which page an answer came from.
To read the manual directly, use the gift icon in the top bar, Product updates and documentation, and choose Documentation. Individual pages also open in a side panel from the question-mark controls beside some page titles and wizard steps, already scrolled to the relevant section.
Getting help from support¶
When something is wrong and rephrasing the question will not fix it, ask Adverity Atlas in the conversation to contact support. It files a ticket with Adverity Support on your behalf, carrying the subject and description you give it along with your organization and the workspace you were in, so nobody has to reconstruct the context. Adverity Atlas confirms the ticket in the conversation and hands back a request ID you can quote in follow-up.
Filing a ticket is available to every role, including Viewer. If your deployment has no support channel configured, the option is not offered and you contact Adverity support by email instead.
Tips for better answers¶
Name the metric, say “ROAS” or “cost per acquisition” instead of “performance”.
Specify the time range, “last 30 days”, “Q2 2026”, or “June 1-15” all work.
Mention the breakdown, “by channel” or “grouped by campaign name” produces a grouped result.
Use follow-ups, each message inherits context.
Reference your data, use @mentions when your workspace has multiple data sources.
Tip
If Adverity Atlas misinterprets a term, the fix is usually in the glossary, not the question. Adding a clear definition solves it for all future conversations.
What’s next?¶
Artifacts, turn answers into mini apps, presentations, and reports
Automations, schedule recurring analyses and alerts
Glossary, improve answer quality
AI providers, change the model behind Chat, or bring your own