# Act on What the Camera Sees

Source: https://help.zira.us/whats-new/vision-hub-training-and-datasets
Summary: Routing rules can now switch an andon light or drive a machine PLC, a trim saw calibrates itself while you watch from your desk, models push to a named slot and prove they landed, and training explains every setting in plain language. You can also connect a site to Claude and ask about your operation out loud, and open the AI Assistant from any page without making it your home page.
Updated: 2026-09-16

Until now, a camera that spotted a problem could tell you about it. In this release it can also do
something about it — a routing rule can switch an andon light or trip a PLC contact the moment it
matches. Alongside that: a trim saw calibrates itself while you watch, models push to a named slot
and tell you they got there, and teaching a camera something new no longer asks you to fill in a
form full of machine-learning vocabulary.

## Ask Claude about your plant

**In beta.** You can now connect a site to **Claude** and ask about your operation the way you would
ask a supervisor:

- *"Why did the trim saw stall last night? Check its readings around 22:00."*
- *"What was yesterday's output on line 3?"*
- *"Show me the open tasks for the Rehovot plant."*

Claude finds the right data source by name and looks the readings up in your live Zira data. No more
copying readings into a chat window or pasting answers back out.

Open your **profile picture** and choose **Connect to Claude**, pick the site, choose what Claude may
do, and paste one line into Claude. That is the whole setup.

Choose **Look only** and Claude can answer questions but change nothing. Choose **Look and act** and
it can open and update tasks, record readings and post to a channel — asking you before each one.

We are still refining this one, and we would like to hear how it goes.

![Connecting a site to Claude: pick the site, choose what Claude may do, press Connect.](/changelogs/vision-hub-training-and-datasets/connect-to-claude/01-connect-panel.png)

Read more: [Connect Zira to Claude](/docs/ai-assistance/connect-to-claude) ·
[the full release note](/whats-new/connect-zira-to-claude)

## The AI Assistant, from any page

Until now, the AI Assistant came with a trade: to use it, it had to become your home page. Plenty of
people would rather start the day in the feed or on a dashboard — so the assistant now has a door of
its own.

- Look for the **AI Assistant** icon — the sparkles — in the left menu, just above **Chats**. It
  opens the same assistant, with the same answers, previews and builders, and your Work mode stays
  exactly as it is.
- **Your saved chats come with you.** A conversation started from the icon is under **Saved chats**
  wherever you open the assistant.
- **On a phone,** open the menu and tap the same icon.
- **Heading back is one click:** open the menu and click the Zira logo to return to your home page.

If the AI Assistant is already your Work mode, nothing changes — it is still the first thing you see,
so the icon stays out of your way.

Read more: [Home AI Assistant](/docs/ai-assistance/home-ai-assistant) ·
[Choose Your Work Mode](/docs/getting-started/profile-preferences/home-page)

## Let a rule switch something on the floor

Routing decides which lane a pallet goes down. Now a rule can also **do something** the moment it
matches — so a decision on the screen becomes an action on the floor.

**Actuators** drive the camera's relay outputs. A relay is a dry contact, so anything that takes a
contact closure can be driven from a rule:

- **Andon lights** — put the red lamp on when a severe defect is routed, and clear it when the line
  recovers.
- **A machine PLC** — stop or start the line, or fire an eject, by closing the contact the PLC is
  already watching.
- Horns, diverters, gates — anything on a relay.

Pick the actuator and relay, then choose **on**, **off**, or **pulse**. A pulse takes a duration,
which is what an eject or a momentary stop signal wants. Actuators that publish named functions of
their own offer those too, with whatever fields the function asks for.

Only actuators the camera currently reports as **online** can be chosen. If a rule points at one
that has gone away, it stays on the rule as a warning rather than being quietly reassigned — so you
find out before it drives the wrong thing.

**Webhooks** are the other half: send an HTTP request when a rule matches, to tell an MES, a
scheduler, or anything else that should know. A rule can call several.

Rules in the **Default** stage can now be set to always match, so pallets nothing else caught can
still trigger an action — not just fall through to a lane.

### And you can tell whether the camera took it

Saving routing sends it to the camera, but a save is not proof it was applied. The toolbar now shows
**Awaiting sync**, **Syncing**, or **Up to date** per camera, with its name when more than one is
paired. Cameras that were offline pick the change up when they reconnect.

## Calibrate a trim saw from your desk

Measuring what a trim saw actually cuts — the length of a board going in and the length coming out —
used to mean commissioning the camera at the machine. It is now a run you start from the console and
watch on the live view.

Open **Trim saw stops** beside the live view, press **Collect**, and let the saw run. The camera
watches boards go past and captures a line for each stop it finds. Stop when you have enough, tell
it which cut size each line is, and apply. That is the calibration.

- **You only declare the sizes.** The positions are the camera's own measurement, so there is
  nothing to tape-measure and nothing to type in from the floor. Declare the sizes up front if you
  know them, or label each line after collecting if you would rather see what turns up.
- **A camera set up against a measured reference needs only one cut size** instead of two. When that
  applies, the window says so and shows the reference it was given.
- **Add to a calibration, or start over — and it asks which.** A saw that has drifted can be
  re-measured while keeping its existing sizes. Clearing them is now a single confirmation naming
  how many stops will go, rather than a question buried under settings for the run afterwards.
- **A run that nobody finished shows up on the data source**, so a collection left open does not
  quietly block the next one.
- Stops are set up by **the side of the board they sit on** — left, right, top or bottom — rather
  than by which end of the travel they are nearer.

This needs the current camera app on the device.

## Send a model to a camera, and know it landed

Cameras can now hold **several models at once**, so pushing one asks which **slot** it goes into.
The slot list comes from the camera itself, so a camera with four of them needs nothing configured
here.

Knowing what a camera is *actually* running used to be guesswork. The Models tab now lists each
model the camera reports, matched against what is stored in Zira, and says how confident that match
is — an exact fingerprint match, or only a likely one based on name and size. Where it is unsure,
you can ask it to **verify**, and it checks the file byte for byte.

Adding a model is easier too: `.pt` files upload directly, and a new dialog collects the type,
classes and where it came from at the point you add it.

## Training that explains itself

There is a **new [Training guide](/docs/ai-vision/training)** in the help centre, covering the whole
path from a labelled dataset to a model running on a camera.

The training form is now called **Train a model**, and it opens by telling you what a run does: it
learns from a dataset, it runs in the cloud, you can close the page, and it takes minutes to hours.

- Settings are grouped into **what you are making** and **how it trains**, so it is obvious which
  handful of fields matter.
- Every field has a plain-language explanation beneath it — *"How many times it reads every image.
  100 is a good start."*
- The expert controls have not gone anywhere; they are collected under **Advanced**, clearly marked
  as things a normal run never needs.
- Only one field is required: the name.

### Fewer ways to waste a run

Cameras need a converted copy of a trained model, and some of the newer conversion formats cannot
run on the camera app yet. If you pick one, the form now says so clearly before you start — instead
of leaving you to discover it when the file will not go to the device.

## Understand how a run went

- Finished runs score **each class separately** on how often the model was right when it fired and
  how much of the real thing it found — so instead of one overall number you can see exactly which
  defect still needs more examples.
- You no longer have to wait for the end. The run's charts now plot **every class as it trains**, so
  you can watch one defect catch up — or fall behind — while there is still time to do something
  about it. Switch between the two scoring styles to tell "found it at all" from "placed it
  accurately".
- The run also lists the **image variations** it generated from your pictures, so you can see what
  the model was trained against. If you need different ones, you can now set them yourself under
  **Advanced**.
- If the run trained on a merged dataset, the job shows **which datasets went into it** and how they
  were split.
- The Training tab says what it is for, and when it is empty it tells you how to start your first run
  rather than showing a bare table.

## Group your datasets

Once a camera has been running for a while, its dataset list gets long. Now you can put datasets into
**groups** — by product line, by camera, by whatever suits you.

- Add a group from a dataset's menu, under **Groups**. A dataset can be in several.
- **The list opens in sections.** A data source with groups is organised, and opening on the
  unorganised version hid that. The plain list is still one click away on the group icon.
- **Rename a group from its heading** and every dataset in it follows. Rename it onto a group that
  already exists and the two merge — you are told before you confirm.
- Datasets that are not in any group collect under **Ungrouped**, after the real groups.
- Anything you put in **archived** drops to the bottom and stays closed, so finished work is out of
  the way without being deleted.

Dataset **tags** have been retired. They did the same job as groups without any of the behaviour, so
a dataset could end up labelled twice and sorted once — groups now do it in one place.

The Datasets table was tidied up while we were there: the toolbar is one consistent set of controls
instead of five different-looking ones, notes show their text in the list rather than an icon, and
the versions count is a link you can click.

## Merging datasets, without the wizard

Combining datasets used to be a three-step wizard ending on a page of counts, duplicate reports,
per-dataset split cards and half a dozen checkboxes. It is now a single short form that asks the four
things a merge actually varies: which datasets, what to call the result, how to split it, and which
classes to keep. Everything else is simply handled.

The detail has not gone — it is in a **Merge analysis** section you can open when a merge looks
wrong, showing per-dataset and per-class counts, repeated images and box quality.

Choosing which classes to keep also works properly now: the list reorders as you press the arrows,
and the number beside each class is the class number the model will learn.

## Follow work that outlives the page

Merges, training runs and camera-file builds take minutes and keep running after you close the
window that started them — or close Zira altogether. **Background work**, the history icon above the
Datasets table, is where you follow them.

It carries a badge with the count, turns into a spinner while something is still going, and opens a
list of what each piece of work is doing and how it ended. The snackbars that start these jobs now
point you straight at it.

## Models that say what they are

- The tab now has **two clearly named lists** — what the selected camera reports it is *Running on
  this camera*, and everything *Stored for this data source*. It used to be one word above both.
- Every model can carry a **type** — what it looks for — which the Models tab groups and filters by.
  You can set it while training, or correct it afterwards from the model's menu.
- **Version** and **Format** are their own columns instead of chips crowded into the name, so long
  model names stop wrapping onto two lines.
- **Opening a training run keeps your place.** It now opens in a panel over the list rather than
  navigating the whole page away from the models you were reading.
- The row menu says what each action is for in plain terms: **Build a file for the camera**,
  **Fine-tune from this model**, **Change what it looks for**.
- Models built in a format the camera app cannot run yet are marked, so you find out in the list
  rather than at the moment you try to send one.

## Quality you can watch live

- **QA snapshot widgets update the moment an image arrives** instead of polling on a timer, so a
  dashboard on the wall keeps pace with the line.
- **A marks report** gathers what reviewers marked across a site, so you can see what is being found
  without opening each session.
- **Auto-defects** flag known defect patterns as they come in.
- A **Validation hub** brings session review together in one tab, and the site list now shows at a
  glance which data sources are stable and current.
- Snapshot widgets take dashboard date filters, so one dashboard can be pointed at any shift.

### Judge a snapshot without leaving the dashboard

Opening a snapshot from a QA Snapshots widget used to load the whole review page inside the
dashboard. It now opens a reviewer built for the job:

- **A magnifier shows the frame at the camera's own resolution** — which is what makes a nail head
  or a hairline crack readable at all. On a pallet camera the reviewer can also draw the board and
  stringer crossings and let you pick the ones a defect is in, and the mark records exactly which
  crops you meant, so the next person looks at the same ones.
- **Your filter comes with you.** The arrows step through the snapshots the widget is already
  showing, so a defect filter carries into the reviewer and keeps going past the end of a capture
  session.
- **Marks already on the image are shown**, so a frame does not get judged twice.
- **Back returns to the dashboard** instead of leaving it, with every other widget still as you
  left it.
- **Or mark without opening anything.** Put the review categories on the card with **Show mark
  buttons**, and on a single-image widget you can draw the crossings there too — tap the ones a
  defect is in, press a category, done. Five steps become two. Opening the snapshot is still the
  better call when the judgement is close, because there the image is full resolution.
- **On a tablet, pinch to zoom** — the magnifier starts off there, since a panel that size costs
  more screen than it gives back. Tap a crossing to select it, and if you do turn the magnifier on,
  press and hold to move it.

## Dashboards and agents

- **Correction fields for Data Research widgets.** Apply a calculation to every reading *before* it
  is grouped, then group, measure or filter by the result — the one place a correction can change
  what a chart's categories actually say.
- **Alerts that get out of the way.** A popup condition alert can now fade itself out after a set
  delay instead of covering the widget until the next reading arrives.
- **AI agents can work across several dashboards** at once, take a time range, and be told what
  answer style you want.
- Charts default to sensible sizes and font scales, gauges to a sensible colour, and a data table
  can be printed.

## More ways to get images in

- **Take a photo straight from your phone** when filling a form or adding images, with the camera
  opening reliably on mobile.
- **Copy into a dataset keeps your choices** — split and annotation settings persist between copies.
- **Shuffle & split** gained real options and a preview of the result before you commit, and both
  copies and merges now honour a **target resolution**.
- Dataset images show **tiny box** and **edge box** chips so weak annotations are easy to spot, and
  label visibility, class colours and thumbnails stay in step.

## On the cameras themselves

Plenty changed on the cameras too. There are no new screens to learn — it is mostly things being
more accurate, and less being lost.

**Sharper pictures, steadier counting.** Exposure, ISO and zoom can be set by hand, the preview can
be rotated and stays that way, and a camera can now sample at high resolution without slowing down
the detection it is running at the same time. A frame-rate default that had been quietly holding
cameras at 15 frames a second is fixed. Occupancy areas can be any shape rather than a rectangle,
counting triggers can be limited to a single label, and pallet cameras are better at nails, lead
boards and keeping detections inside the pallet — with calibration that stays correct when the
camera is mounted at an angle. Stations running several cameras together coordinate more reliably.

**Less gets lost.** Shift counts now survive restarts, recoveries and shift changes instead of being
dropped, and a reading the server rejects is held rather than discarded. Recordings keep their raw
and validation clips together, pick up stalled uploads again, and recover from encoding errors
rather than ending up half sent. Cameras also show their CPU and battery temperature, so a unit
working too hard is visible before it starts dropping frames.

## Around the rest of the app

- **Image review** — a new **Pallet mark** filter narrows a session to one product or batch. It shows
  up only when the session actually contains marked pallets.
- **Pie charts** can now show a **legend**, for when the slice labels are too tight to read.
- **Indicator widgets** read **True** or **False** on a yes/no field instead of a raw value.
- **Chart tooltips** hold still, so the **view entries** button stops sliding away as you reach for
  it — and sorted Data Research charts now draw their axis correctly, embedded ones included.
- **The tasks board** scrolls to the bottom, loads more reliably, and opens the card you clicked.
- **Schedules** can be added at site level, and straight from an AI widget.
- The app **remembers more between visits**, so returning to a page you were just on is quicker.

## Read more

- [Connect Zira to Claude](/docs/ai-assistance/connect-to-claude) — new guide
- [Home AI Assistant](/docs/ai-assistance/home-ai-assistant) — opening it from the left menu
- [Routing](/docs/ai-vision/routing) — actuators, webhooks and device sync
- [Trim Saw Stops](/docs/ai-vision/trim-saw-stops) — calibrating a trim saw
- [Models](/docs/ai-vision/models) — rewritten
- [Training](/docs/ai-vision/training) — new guide
- [Datasets](/docs/ai-vision/datasets) — groups, merging and background work
- [QA Snapshots](/docs/dashboards/qa-snapshots-widget) — live updates, marks and reviewing from the dashboard
- [Image Review](/docs/data-sources/image-review) — pallet mark filter