08Documentation
Reference
Look it up: the tools, the caps, the models, the words.
Supported tools
The 18 tools Vernais connects to, what each one brings in, and which four carry the event data a why question needs.
The 18 tools
If your tool is in this table, Vernais can read it. If it is not, there is no way to add it today.
The catalog is fixed at 18 tools and 72 streams. A stream is one kind of record a tool hands over — Stripe has charges, refunds and invoices; Jira has issues, projects and users. You turn each stream on or off yourself.
The four tools that carry event telemetry are listed first. The rest follow in alphabetical order.
| Tool | What it brings in | Event telemetry? |
|---|---|---|
| Amplitude | Product events, cohorts | Yes |
| Mixpanel | Product events, user profiles | Yes |
| PostHog | Product events, insights, persons | Yes |
| Stripe | Charges, refunds, invoices, payment intents, subscriptions, customers, products | Yes |
| Confluence | Spaces, pages, blog posts, attachments | No |
| Datadog | Events, incidents, monitors, dashboards | No |
| Figma | Team projects, project files, file comments, your account | No |
| GitHub | Repositories, commits, issues, pull requests, releases, workflow runs | No |
| GitLab | Projects, issues, merge requests, pipelines, users | No |
| Google Drive | Drives, files | No |
| Intercom | Conversations, contacts, companies, admins, tags | No |
| Jira | Issues, projects, users | No |
| Linear | Issues, projects, teams, comments, users | No |
| MS Teams | Teams, channels, channel messages, users | No |
| Notion | Databases, pages, users | No |
| Sentry | Issues, releases, projects, organizations | No |
| Slack | Channels, messages, files, users, team info | No |
| Zendesk | Tickets, users, organizations, groups | No |
Each tool asks for its own credentials. Stripe wants one secret key. Jira wants a subdomain, an email and an API token. The connect form shows only the fields that tool needs, and it ships with a How to get these credentials guide. See Connect a tool.
What the telemetry column means
Event telemetry is a record of one thing happening to one user at one moment — a purchase, a failed checkout, a charge.
Four tools carry it: Amplitude, Mixpanel, PostHog and Stripe. That is what lets Vernais answer a checkout question with a success-versus-failure split — how many went through, how many did not — instead of "there are some tickets about checkout".
The tool's name is not what decides this. Vernais opens each record and checks it. Does the record have a specific event name, not a generic one like event or log? Can a real time be read out of its own fields? Does it come from a product-analytics tool, or is it a payment with a transaction shape? A record that fails any of those checks is not counted as an event.
Your other tools still count for plenty. A Sentry issue, a Jira ticket, a Slack thread — these are evidence, and an investigation quotes them by name. They are not counted in the success-versus-failure math. The code says it plainly: a record can be qualitative evidence without being valid business telemetry.
This is why naming the right tool matters when you ask why something is happening. Name Mixpanel or Stripe and you get numbers moving over time. Name only Jira and you get the discussion about the numbers. See Find a root cause.
Skip the telemetry tools when your question has no number in it. "What did we decide about the refund policy?" lives in Slack and Notion, and Stripe has nothing to say about it. The test: can you picture a number moving? If not, you want records, not events.
Words that mean more than one tool
Some everyday plurals fan out to a fixed set of tools, so a question you thought was about one tool quietly covers three.
| You type | Vernais reads it as |
|---|---|
| tickets | Zendesk, Intercom, Jira |
| errors | Sentry, Datadog |
| exceptions | Sentry |
| docs | Notion, Confluence, Google Drive |
| analytics | Amplitude, Mixpanel, PostHog |
| PRs, pull requests | GitHub, GitLab |
| messages | Slack, MS Teams |
| charges, payments | Stripe |
| designs | Figma |
"How many tickets do we have?" is a three-tool question. If you meant one, name it. The fan-out is trimmed to the tools that actually hold data in your workspace, so a group word covering a tool you never connected quietly narrows to the ones you did. More on phrasing in Ask a data question.
What one sync brings in
A sync is a capped pull, not a full history download.
doc may never carry a valueThe field tree in the Field mapping tab merges two sources. Fields found in your own stored records show a fill-rate bar. Fields the vendor documents but that have never appeared in your data show a doc badge instead. Before a first sync you see only the documented set, and for some streams that set is empty.
Syncing stores records. It does not make them answerable. The knowledge graph is a separate build, and the first one is manual. See Build the graph.
Streams that need a parent record
A few streams hang off a parent — a repo, a project, a team, a channel or a file. Vernais cannot pick that parent for you yet, so these streams do not pull from a live connection today.
- GitHub — commits, releases, workflow runs (each needs a repository)
- GitLab — pipelines (needs a project)
- Slack — messages (needs a channel)
- MS Teams — channels, channel messages (need a team or a channel)
- Figma — team projects, project files, file comments (need a team, project or file)
The rest of the run still succeeds. Open the Logs tab and expand a run to see which single stream stopped and why.
The catalog is shared, your data is not
The same 18 cards appear in every workspace; everything behind them is per-workspace.
Credentials, stream choices, field choices and every synced record live inside one workspace's own database. Open a new workspace and you see the same 18 tools with nothing connected and zero records. Connecting Stripe in one workspace does not connect it in another.
Integrations is an elevated section and it is off for new people by default. The workspace owner grants it per person. Without that grant, a teammate does not see Integrations in the nav at all. See Permissions and seats.
Questions we get
Our tool isn't on the list. Can you add it?
Sentry holds all our errors. Why isn't that event telemetry?
Do I have to turn on every stream?
Related resources
Limits and caps
Every cap in one place: the number, the boundary that resets it, and what to do instead.
One constraint shapes most of what follows: when Vernais answers from your connected data, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.
That is why the caps below exist, and why almost none of them hide anything from you. When a cap bites, the answer says so on its face. Link to this page from your own notes instead of copying the numbers — numbers move, and a copied number rots quietly.
Answers in chat
Ask for a count when you want a total. Never add up the rows on your screen.
issues about promo codes) ranks by meaning against a recall pool first, then filters it. So that total covers the pool. A search filtered by tool, stream or field skips the pool and its total covers every record.(+N more across M other values), and records missing that field are counted on their own line.N events beyond the top values went unclassified.scan capped ... count may be a lower bound, treat the number as a floor. Name the tool or the stream and ask again.Every number arrives with the filter that produced it, like count = 200 matching [integrations=['stripe']]. Read that recipe before you trust the number. A right-looking answer to the wrong question looks exactly like a right answer.
The counter-case: if you only want to eyeball a handful of real examples, the list cap costs you nothing. The test — are you about to count what you see, or read what you see?
Investigations
A root-cause run takes a few minutes. Give it a narrow subject and it has less ground to cover.
Web search, deep research, and scraping
Scraping saves pages for later. Searching answers you now and keeps nothing. Pick the one that matches what you want to still have tomorrow.
All pages is not infinite — it clamps to 50, and any number you type is bounded to 50 as well.Stop in the Web scrape panel. Everything collected so far is kept and stays searchable.AI Web scraping and send your topic again. It restarts from the first question.Connecting and syncing a tool
A sync is a top-up, not a full history download. Run it more often rather than expecting one run to fetch everything.
Settings tab. The window then rolls forward on its own.doc badge come from that sample, not from every record you hold.Logs tab shows exactly when it did.Inbound webhooks
Batch your events into fewer, fuller deliveries. Every cap here rewards that.
Files, pages, and tagging
If a file is too big, the part you care about is usually a page or two of it. Paste that.
file too large with the size it saw.Seats, workspaces, and people
A seat is a billing ceiling. Permissions live inside it — they cannot lift it.
This workspace is full.Power seat can work with data but never chat or investigations, and no per-person grant can buy that back.Max seat if they need to ask the AI anything.The counter-case for seats: someone who only reads dashboards and records does not need a Max seat. The test — will this person ever ask a question in chat?
Related resources
Models and thinking
Three models can answer your chat, and one switch changes how they think. Only one setting on that switch makes a difference you can feel.
One constraint shapes most of what follows: when Vernais answers from your connected data, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.
The model is the writer, not the source. Swapping it does not change which records exist, or what the tools return. It changes who phrases the reply.
The models you can pick
Open the chat bar and click the model button — it shows the model you are on now. Three models are on the list.
| Model | Runs on | Cost | Context window |
|---|---|---|---|
| DeepSeek V4 Flash | Fireworks | Paid | About 1M tokens |
| Nemotron 3 Ultra 550B | OpenRouter | Free tier | About 1M tokens |
| Nemotron 3 Super 120B | OpenRouter | Free tier | About 1M tokens |
All three call tools natively. All three can have thinking turned off. Your pick applies to every turn, including code questions, and it is remembered across sessions.
Leave the default alone unless a reply fails or you need one very long answer.
Why DeepSeek V4 Flash is the default
It is the only paid model on the list. The other two are free tiers, on a different provider. A paid primary with free fallbacks means trouble at one provider rarely stops your turn.
It is also the model the rest of Vernais runs on. The investigator, the topic namer and the writing passes all share one chain, and DeepSeek heads it.
Pick Nemotron 3 Super when you want a single very long answer. It has the most room per reply of the three — about 262,000 tokens, against about 131,000 for DeepSeek and about 66,000 for Ultra. The test: if a long write-up keeps getting cut short, the room per reply is the thing to change.
When another model answers instead
Your pick is tried first. If it errors before writing a single word, the server walks a fixed chain: DeepSeek V4 Flash, then Nemotron 3 Ultra, then Nemotron 3 Super.
A small italic note under the reply names the model that actually answered. Read it when a reply feels different in tone or depth — that is usually the explanation.
The fallback only fires before any content is written. Once a model has started the answer, Vernais will not splice a second model into it. A reply cut short by a failure says so instead.
The thinking switch
Open the model menu and scroll past the model list. At the bottom is a Thinking section, headed by an On/Off switch. It is on by default for all three models.
Off means exactly zero reasoning. The model answers straight from your message and the tool results. Nothing is thought through first, and nothing appears in the thinking block.
On means the model reasons before it answers. You see that reasoning above the reply, interleaved with each tool step.
Turn thinking off when you want a fast count and you already know which tool it needs.
Effort levels
When thinking is On, an effort control appears under the switch. It disappears when thinking is Off, so the menu never offers a dial that cannot apply.
| Model | Effort rungs | Can turn off |
|---|---|---|
| DeepSeek V4 Flash | low · medium · high · xhigh · max | Yes |
| Nemotron 3 Ultra 550B | minimal · low · medium · high · xhigh · max | Yes |
| Nemotron 3 Super 120B | minimal · low · medium · high · xhigh · max | Yes |
Until you pick a rung, the menu shows High for all three. DeepSeek has no minimal rung because its provider rejects that value outright.
What the rungs actually do
We measured this rather than guessed it. One fixed prompt, run through DeepSeek at each rung, produced this much reasoning:
Read that list again. low, medium and high sit within noise of each other, and the order is not even reliable — medium came out shorter than low. Only the ends move. xhigh and max climb, and Off is exactly zero.
Off is the one reasoning control that reliably changes what you get. Do not expect to feel a difference between low, medium and high. We looked for it, and it is not there.
The two Nemotrons work differently underneath. Their provider accepts an effort word but ignores it. So Vernais converts your rung into a thinking budget in tokens instead, and the model stops under that budget. The low rungs bite there; the high ones leave it free.
What your model choice does not change
- The investigation — a why-question runs its own pipeline with its own pinned models. It always reasons hard. Your chat-bar choice never reaches it.
- The jury — inside an investigation, the jurors that score evidence run on a separate small model. Nothing in the chat bar touches them.
- Deep research — the
Deep researchtoggle in the + menu forces thinking to high, whatever your switch says. - The rest of Vernais — writing an initiative, naming a topic, scoring a fit: all pinned to maximum reasoning, outside your control.
- Grounding — no model may answer a question about your data without running a tool first. Turning thinking off does not turn that rule off.
Limits and resets
minimal rung on DeepSeeklow, or switch to a Nemotron if you want minimalRelated resources
Glossary
Twenty words Vernais uses that do not mean what you would guess. One plain sentence each.
One constraint shapes most of what follows: when Vernais answers from your connected data, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.
The map: what Vernais learns from
Everything the AI tells you traces back to a signal you can open and read for yourself.
A map of your company's data: every record is a dot, and every provable link between two records is a line.
One record from one connected tool, cleaned and de-duplicated — a single Jira issue, one Slack message, one Stripe charge.
A real thing that many records talk about, like a customer or a version number, with every different wording for it merged into one.
A group of records that mean similar things, found by maths and named with a real phrase lifted from the records themselves.
Open a record and you see the tool's own fields plus a Raw payload section holding the original, untouched. Open an entity and you see a Resolved from list: the exact wordings that were merged into it.
Getting data in
Syncing records is not the same as building the map — the brain is a separate run, and nothing appears in Data until it finishes.
One pull of new records from a connected tool into Vernais.
One feed inside a tool — Stripe has charges, Jira has issues.
One named piece of data inside a record, like a Jira ticket's priority, and you choose which ones Vernais pulls.
A bookmark: the value marking how far the last sync got, so the next sync pulls only newer records.
Pages Vernais read from the public web, listed under their own Scrapes tab rather than mixed in with your connected tools.
The run that turns your synced records into the knowledge graph, started from Integrations → Run Corvex.
The brain runs 15 stages and takes a few minutes. It needs an elevated permission, and it builds only into the workspace you are in — so a new workspace stays empty until someone runs it there.
The answer
An investigation's conclusion is never saved as a fact. Vernais works the cause out fresh each time you ask, and hands you the evidence with it.
A candidate explanation for why your problem is happening, carrying the records that support it.
How sure Vernais is that it found the real cause: high, medium, low, or hypothesis only.
One number, with its formula shown, that measures the exact problem the investigation just diagnosed.
A causal answer can reach high confidence. The model proposes a label and later checks can only pull it down, never push it up. "Hypothesis only" means Vernais found a plausible mechanism but refused to call it a proven cause.
The work
An initiative born from an investigation arrives full; one you create by hand starts with only its title.
One piece of product work with a problem at its centre, usually built from a finished investigation.
A named group of an initiative's evidence records, so you read a few themes instead of hundreds of rows.
The tracking metric's value at the moment you lock it, frozen so your "before" number cannot drift.
The container you file work under, so one product's investigations, initiatives and scrapes sit together.
Picking a product tags the work it produces. It does not narrow what an investigation searches — a question still runs across everything the graph holds.
Access
A seat sets the ceiling, and permissions work inside it.
One company's private container, with its own connected data, its own graph and its own initiatives.
What a person is billed for: a Max seat has full AI, a Power seat has data only and no AI.
One switch the workspace owner turns on or off for one person, like "run Corvex".
There are no roles here. The owner grants each person their own set of switches. Basic viewing rights start on, the manage-and-admin ones start off, and the two AI rights need a Max seat — no switch can hand AI to a Power seat.
Pairs people mix up
| Term | Mistaken for | The difference |
|---|---|---|
| Signal | Signal cluster | A signal is one record; a cluster is a named group of them. |
| Topic | Signal cluster | Vernais builds topics across all your data; a cluster groups one initiative's evidence. |
| Sync | The brain | A sync brings records in; the brain turns them into the map. |
| Tracking metric | Baseline | The metric is the formula; the baseline is its frozen value at lock time. |
| Seat | Permission | A seat is what you pay for and it caps AI; a permission is one switch inside that cap. |