Ask a question
Step by step, from phrasing a question to keeping the answer as a tracker, report or board card.
This is the full workflow for getting an answer you can use. For a first look, see Ask your first question.
Check you are in the right project
The workspace switcher at the top of the sidebar shows the current project. Answers only use that project's sources and definitions.
Open Reasoning
Sidebar → Reasoning.

A new conversation starts with What you can ask: example questions your analytics team keeps for this project, grouped by topic. Click one to ask it. Not available yet lists topics the team does not answer yet and why; asking about one of them gets that reason back instead of an answer, with questions you can ask instead.
You can also ask Tattva itself: What can you answer? lists, for this project, the example questions, the metrics by area (with the date each is available from), the data and how fresh it is, topics not answered yet, question kinds switched off, how figures are made and what your personal-data access lets you see. Do you do ROAS? (or any metric or topic) gets a one-line yes, no with the team's reason, or partly with the date it is available from, plus questions to try. Click a suggested question to ask it.
To continue an earlier conversation, click Sessions and pick it. To start fresh, click New in the Sessions drawer.

The drawer lists your own sessions, most recent activity first. Scroll down to load older ones. To find one:
- Search matches the session's title, every question in it, and the first line of each answer (its headline). Words can be in any order, and part of a word works too ("revenu" finds "net revenue"). Put a phrase in quotes to match it exactly, or put
-before a word to leave out sessions that have it. While you search, each session shows the question or answer line that matched. - This project lists the current project's sessions. Pick another project, or All my projects, to see sessions you asked in other projects you belong to. Opening one switches to its project.
- Any time narrows the list by when a session was last active: Today, Last 7 days, Last 30 days, or Dates… for a range (both days included).
- Me / Everyone / a person lists another person's sessions, or everyone's. A superadmin can pick anyone, in any project. A workspace owner or admin can pick the people with sessions in that workspace's projects, and sees only those projects' sessions; if the project filter is set to a project outside the workspaces you administer, picking a person switches it to All my admin projects. Everyone else sees only their own sessions in the drawer. A link to a session still opens for any member of its project.
Phrase the question
Plain English is fine. A clear question gets a faster, cleaner answer.
Patterns that work:
- "What was metric in period for segment?": a value.
- "How did metric move week over week for the last 8 weeks?": a trend.
- "Why did metric change between period A and period B?": a root cause.
- "Forecast metric for next month.": a forecast.
- "Which of A, B, C contributed most to metric?": attribution.
Things that help:
- Name the period. "Last month" or "1–31 March", not "recently".
- Name values you care about. "Net revenue from Warehouse 1" lets Tattva fill in the warehouse without asking.
- One question at a time. Split "X, and also Y, and explain Z" into three.
- Be explicit. "Compared with February", not "compared with that".
- Top or bottom. "Bottom 5 products by net revenue" lists the lowest five.
- Excluding. "Net revenue excluding Dia Free Juice" gives every other product's total. If no definition can leave it out, Tattva asks.
- Week on week. "WoW net revenue" compares the last full week with the week before. Say "this week so far" for the week that is not over yet.
- A skill by name. "Top 5 products last month using Coverage movers" has the planner follow that uploaded skill's playbook. If it isn't offered for the question, or no skill has that name, the answer says so and answers the rest. See Naming a skill in a question.
- How you want the answer. The figures are worked out first and never change with these; then Tattva shows them as you ask, using only its own tables, columns and rows:
- Language: "answer in Hindi" answers in Hindi (Devanagari), or in Roman-letter Hindi or English when you ask for those, whatever language the question is written in.
- A chart: "as a waterfall", "as a line chart", "as KPI cards" or "as a graph" draws it under the table. A waterfall is drawn only when the rows add up to a total Tattva computed (a P&L's lines, each product's CM3 adding to the total); otherwise the answer keeps its usual layout and says why. A chart type it does not draw ("as a pie") is named as not shown; pick it from the chart's Chart type list.
- Order, columns and rows: "lowest first", "only the product and CM3 columns", "show only 5 rows", "all rows". A cut keeps the Total, labelled Total (all N).
- Just the number: "just the number" gives the headline figure without the table.
- Definitions only: "from the stored definitions only" answers this question only from the team's stored definitions: no query of the agent's own, and a skill that writes its own queries is not used.
Send
Press Enter. (Shift + Enter adds a new line.) Watch the work in the Trace rail: Analyze always, then the answer is written and checked. A question that needs several separate results (different periods for different products, the difference between two changes) is planned instead: the rail shows Plan · N steps, and each step appears as it runs. A planned answer took a median of 55 seconds on prod before the claim check, and 131 seconds when it follows a playbook; the claim check's read, revision and re-read add about 39 seconds (measured on 16 recorded questions). Click Stop if you want to halt the run.
Answer any clarification
If your question can be read more than one way, or a metric needs a value you didn't give, Tattva stops and shows a Needs clarification card.

Click one of the choices, or type in "Or type your own answer…" and click Reply. Tattva asks one thing at a time; you may get a second card if two things were missing.
Read the answer, then the Trail
The Answer tab has the answer; see Reading an answer below. The Trail · N tab has the evidence:
- How this was answered comes first. It lists only what changed against what you asked, each with a status: Adjusted (Tattva changed it, for example a day still loading), Assumed (your words could mean more than one thing and Tattva picked one), Corrected (you corrected yourself in the question), Not applied (part of the question is not in the answer), Switched off (your project doesn't answer this kind of question), Asked you, Not what happened (the figures went the other way from what the question assumed) or Reused. When nothing changed it says "Answered as asked." A follow-up starts with Read as: and your question written out whole, with the parts kept from earlier questions marked, so you can see at a glance whether Tattva followed the conversation.
- Details (closed) lists every part of the question, the period and where each table's Total came from. Anyone can open it.
- Debug (closed) shows how the model's reading compared with code's. Only superadmins and project owners, admins and ontology admins see it.
- Then, for most answers, the phase cards: on the Analyze card, the Compute block has each metric's value, filters, inputs and SQL.
- A planned answer shows Steps instead of phase cards: one line per step in plain words with its result ("Ran Product Net Revenue · DIA FREE JUICE · Aug vs Sep 2026 → ₹3.60 crore → ₹3.54 crore"), and How each figure was derived. Click a figure in the answer to jump to the steps it came from.
A sentence of the prose that failed the claim check was revised once or dropped before you saw the answer. On answers recorded before the claim check, a grey note after the answer, "An earlier draft was checked and revised before this answer was shown", means the self-check rejected a first draft; See what was run opens the withdrawn attempt and Why? lists the reasons.
For how to check an answer properly, see Review an answer.
If your team has registered a dashboard that covers your question, you may get a link to it instead of a full answer: the dashboard's title as a link, what it shows, when it refreshes, and the note "Built by the team; its figures may differ from Tattva's definitions." Click Answer here instead to get the full answer from Tattva. If there is no button, your project is set to Link only; ask your admin if you need the full answer. Questions about why something changed, forecasts, what-if questions, definitions and comparisons of two periods always get a full answer.
If someone asked the same thing recently, you may get their answer straight away, starting with "Answered on 18 Oct 2026; the data has not changed since." This happens only when your project's Repeat questions setting is on and all of these hold:
- the earlier question resolved to exactly the same metrics, dates, breakdown and filters (the wording does not matter, but "last 15 days" asked on another day is a different period);
- it was answered in the last 30 days, and no metric definition (its query or its description), convention, settlement, family policy or available-from date it used has changed since, nor the project's Figures from setting or word meanings, nor (for an answer from an uploaded skill) the skill's version, and no Tattva update since has changed how answers are worked out;
- none of the warehouse tables it read has been loaded since it read them, and it read nothing whose changes Tattva cannot check (an uploaded file, a document, the web);
- the person who asked it had the same personal-data access as you;
- the earlier answer passed its checks cleanly (no warning, not revised, nothing removed), and nobody marked it Not satisfied.
So a stored answer lasts at most 30 days, and in practice only until the next load of any table it read. Tables that load several times a day keep it for hours, not days.
Click Ask again under the answer to have Tattva answer it in full now. If the stored answer is wrong or out of date, an owner, admin or ontology admin can click Clear stored answer under it, so that nobody gets it again and the next person to ask gets a full answer.
Keep what is useful
The toolbar under the answer:

| Button | What it does |
|---|---|
| Save as report | Creates a report from the answer and opens it. See Reports. |
| Save as tracker | Creates a daily tracker with this question and opens it. Add a rule there. See Trackers. |
| Copy tables (TSV) | Copies the answer's tables, ready to paste into a spreadsheet. Figures paste as plain numbers (no ₹, %, or digit grouping, - for negatives), so Excel and Sheets can add them up; the unit is in the column's heading ("Net Revenue (₹)"), and a first row names the table, its dates and its filters. A long table shown cut ("Other 77") pastes every row. Lone dashes paste as empty cells. Pasted into Excel or Sheets, bold total rows stay bold. Personal data redacted from the answer is redacted in the copy too. Shown only when the answer has tables. |
| Pin to board | Adds the answer to a board as a note card. See Boards. |
Tell Tattva how it did
Under the answer, Was this answer what you needed? has Satisfied (thumbs-up) and Not satisfied (thumbs-down).
Satisfied is saved with the answer and the button changes to Thanks!.
Not satisfied opens What wasn't right?. Pick any of the quick reasons: Figure looks wrong, Definition is not what we use, Wrong period, Breakdown missing, Wording unclear or Other. Then say what was wrong and, if you know it, what is correct. Pick at least one reason or say what was wrong, then click Send feedback. The dialog thanks you. Your comment goes to the project's analytics team (owners, admins and ontology admins) with the question, the answer and its Trail. It does not change any answer, not even yours, until the team settles it; the team makes the fix in the ontology so every member gets the same answer.
After you send it, the dialog offers Re-ask with this correction. It sends a follow-up question in the same conversation that asks Tattva to answer again taking your note into account. That is the only immediate effect, and only in that conversation.
Under the answer, a line shows the status, for example "You flagged this answer — Pending DA review", "— In review by the team", "— Applied as convention “…”", "— Already covered by stored definition “…”" or "— Not applied: …", with a My corrections link. You get a notification when your comment is settled. All your comments are listed in Ontology › Maintain › Review › My corrections.
Follow up
Ask the next question in the same session. Tattva keeps the conversation's last five questions and answers, including what each was checked as and the rows of each table: "and in September?" keeps the metric and the product, and "the second one" points at the second row of the last answer's table. A part you don't restate comes from the latest question that had it. If a filter from an earlier question is not applied, the answer says so. Check Read as: at the top of the Trail to see how the follow-up was read. You can reopen the session later from Sessions.
Reading an answer
An answer about your metrics is laid out the same way every time, whether it was answered in one run or planned:
- The verdict sentence gives the figure you asked for, and the change when you compared periods. If your question assumed a rise or a fall and the figures say otherwise, this sentence says so first ("Net Revenue did not fall: it rose 8.69%"). A model writes it and the rest of the prose; a second model lists every claim they make and code checks each against the figures. A sentence that fails is revised once, then dropped; when the verdict sentence goes, code's own verdict line takes its place.
- The rest of the prose: what the figures show, in the model's words, every number and comparison checked.
- Tables and a chart: the figures in full, with totals only where the rows really add up. A table of products shows the largest first.
- Notes under the tables (each starting with ⓘ): anything that changes how you read the figure, such as a day still loading, words in your question that were not used, or a figure that will still change (Provisional or Incomplete).
- The Period line names the exact dates the figures cover, with weekdays for a week or less. "This month vs last month" means the last complete month against the one before; say "so far" if you want the month in progress.
- The Scope line says what data the answer covers, and the Source line which definitions and tables it used.
If something looks off, open Trail and read How this was answered: it lists what Tattva changed or assumed. If the figure is still wrong, click Not satisfied and say why.
Common pitfalls
- Wrong project. Check the switcher before you ask.
- Missing data. If the answer says the data isn't there, it usually isn't. Connect the source, or ask your data analyst.
- Unclear period. "Recently" means different things on different days.
- Ignoring provisional figures. A Compute entry marked Provisional or Incomplete will change. Say so if you share it.
- "The data warehouse was busy." BigQuery stopped a query at its two-minute limit, usually because it waited for capacity. The answer says how long it queued and ran, and nothing was computed. Ask again in a few minutes.
Asking from Manas or Slack
You can ask the same questions from Manas, in its web chat or in Slack. See Use Tattva from Manas.