Troubleshooting
Common problems in Tattva and what to do about each.
Signing in
I keep landing back on the sign-in page
Your session has ended (sessions last 30 days) or your browser dropped the cookie. Sign in again with Continue with Google. If it keeps happening within minutes:
- Check that your browser allows cookies for the Tattva site. Blocking third-party cookies can break Google sign-in.
- Try a private window. If that works, a browser extension is interfering.
- If sign-in itself loops, send your operator the URL and a screenshot of the browser console.
I can't ask questions
Viewers can read projects but can't ask questions. Ask an Admin or Owner to make you a Member. See Permissions.
Sources
My file never reaches Ready
Open Sources and select the file.
- "Pipeline failed — check logs." The file couldn't be read. Common causes: a malformed header row, a truncated file, an encoding other than UTF-8, or an empty sheet. Fix the file and upload it again with the same name; it becomes a new version.
- It's a PDF. PDF uploads are refused. Save the text as
.mdor.txt. - It's
.jsonor.parquet. These go through the spreadsheet reader. Convert to CSV, or load the data into BigQuery. - It's still processing. Large files take a while. If a small file sits in Parsing for more than a few minutes, refresh the page. If it doesn't move, send the source's name to your operator.
My metric document didn't produce anything to review
- The document must be
.mdor.txt. HTML files are stored but not read as metric documents. - Documents are read by the ontology build. Run one with Regenerate on the Ontology page.
- Watch the progress panel while it runs. Under the
rules_extractionphase, any ⚠ line tells you what part of the document could not be read and why. - Then check Ontology › Maintain › Review for new candidates.
Answers
The answer looks wrong
Don't re-ask the same question hoping for a different number. Find out where the number came from:
- Open the Trail tab on the answer. Each phase is listed, with every attempt.
- Open the Compute card for the metric. It shows the value, the Filters, the Parameter / Value / Source table (which dates and values were used, and where each came from), any Joins, and the SQL. Most wrong answers are a wrong input: the wrong period, the wrong channel, a value that didn't match.
- Check the label. A figure computed from a stored definition comes with its Compute card. A figure marked as the agent's own SQL wasn't checked against a definition; ask your data analyst to define that metric.
- Check the definition. If the SQL itself is wrong, the metric's definition needs fixing. Your data analyst can open the metric's Definition tab, compare with Try it, and correct it. See Metrics in Make.
- Check for a family. If the metric is one of several versions (for example several kinds of revenue), the answer may have used a different version than you meant. Look for the Metric family block in the Trail and ask again naming the version you want. See Families.
Then click Not satisfied under the answer, pick the reasons and say what was wrong. Your comment goes to the analytics team, who fix the ontology so everyone gets the right answer; it changes no answer until they do. To get a new answer straight away, click Re-ask with this correction after sending.
The answer says "Provisional" or "Incomplete"
That's intended, not an error.
- Provisional: the period includes recent days whose figures haven't settled (for example, returns aren't known yet). Settled days use actual figures and the rest is estimated. The figure will change.
- Incomplete: the data needed for this figure doesn't exist yet, for example a cohort whose outcome window hasn't closed. The figure will change.
See How answers work.
An answer was "withdrawn by the self-audit"
Tattva checks its own answer before showing it. When a figure, a sum, a completeness check or one of your project's conventions failed, the first attempt was withdrawn and written again. A grey note after the answer reads "An earlier draft was checked and revised before this answer was shown"; click Why? for the reason, or See what was run to open the Trail at the withdrawn attempt. The answer you see is the revised one.
Tattva asks me a question back
A Needs clarification card means one input was missing and couldn't be worked out from your question, such as which channel or which version of a metric. Click one of the choices, or type your own answer in Or type your own answer… and press Enter. Tattva asks for one thing at a time.
To avoid it next time, name the value in the question: "net revenue from Warehouse 1 last month".
Tattva says it doesn't have the data
- The source isn't connected, or isn't Ready. Check the Sources page.
- The metric isn't defined. Ask your data analyst to add it.
- The data is personal and your PII clearance doesn't cover it. Ask an Admin. See Sensitive data.
Tattva says it doesn't answer that yet
Your analytics team decides what Tattva answers in each project. The reply says which rule applied:
- The topic is not answered yet. The team marked it (for example a metric whose data is not loaded), and the reply gives their reason. Pick one of the suggested questions instead, or tell the team you need it.
- That kind of question is switched off, for example forecasts or "what if". Ask for the figures themselves ("net revenue by month this year") instead.
- "Not available before …" The metric's figures start on that date. The rest of your period is answered.
- A figure shows "—" with "not defined yet". The project only shows figures from metrics the team has defined, and this one has no definition yet. Ask your data analyst to define it.
Your question is still recorded, so the team sees what people ask for. What you can ask on an empty Reasoning page lists questions that work. See Answer scope.
"Continue in chat" opened an empty composer
From a Triage card or a tracker question, Continue in chat opens Reasoning without your question filled in. Copy the question from the card and paste it. The Continue in chat → link on a tracker's History tab does fill it in.
Ontology
I edited a column's meaning but answers ignore it
Your edit is saved but not yet in effect. The Ontology page shows a banner saying how many curation edits are "saved but not yet in effect". Run a build (click Rebuild on the banner, or Regenerate) and ask again.
Accept is greyed out on a candidate
A blocking check failed. Hover over Accept to see why, and read the checklist on the card: V1–V4 always block when they fail.
A document that has already produced proposals is not read again by a later build. To change a definition, upload the corrected section as a new document, click Regenerate, then Reject the old card. To have the same document read afresh instead, click Read again on next build on its source in Sources, then Regenerate. See Review what Tattva proposes.
Trackers
A tracker question shows "Not run yet"
It has never run. Either the tracker is on Manual / on demand (paused trackers are set to manual), or its first scheduled run hasn't come round yet. Click Run now on the tracker to run every question now. To schedule it, click Edit on the question and pick hourly, daily or weekly.
A tracker question shows "Skipped"
The run happened but couldn't compute the question. The reason is shown next to it as "Skipped —" followed by the reason. The usual ones:
- "Couldn't infer a metric from the question." Name the metric in the question, or set a rule that names its
metric_slug. - "Metric … not found in this project's ontology." The metric isn't defined in this project. Ask your data analyst to define it, or edit the question to use one that exists.
- "Couldn't compute …" The metric exists but couldn't be run, and the rest of the message says why.
Fix the cause, then click Run now.
A question always reads the same and never flags anything
Look for the amber no rule chip. Without a rule (or a linked goal), runs are never classified and the question shows All clear. Click Edit on the question and set the Classification rule to Relative delta: it is the rule that checks the measured change against your critical if Δ > percentage. An LLM judge rule is saved but not applied, so it also shows All clear; Absolute threshold and Hypothesis invalidated mark every run Needs attention. See Trackers.
Run now finishes but nothing appears
The message "Run completed but no rows were created" means the questions reference metrics that don't exist in the active project. Check you're in the right project, and that the metric is defined.
I want a tracker to run monthly, or at a set time
Trackers run hourly, daily or weekly, or on demand. There's no monthly cadence and no time-of-day setting. For a monthly check, use a weekly tracker, or run it on demand with Run now when you need it.
Settings
My API key isn't accepted
- Send it as
Authorization: Bearer <token>. - Check on Settings › API keys that it isn't marked revoked or expired.
- A key has its creator's permissions. If their role was lowered, the key can do less too.
I can't open the audit log
Only workspace owners can. Ask an owner.
Still stuck
- Note the URL, the time and what you were doing.
- Open your browser's developer tools and copy any red errors from the console.
- Take a screenshot.
- Send all three to your operator's support channel.