Skip to main content

Catalogue Health Dashboard

The Catalogue Health Dashboard tells you which parts of your catalogue are hurting your recognition accuracy, and why. Rather than reporting on what was detected in your images, it reports on the catalogue itself — the reference images and variations the models match against.

This matters because most accuracy problems are catalogue problems. When two products are consistently confused, the cause is usually that their reference sets overlap, or that a bad crop was approved into one of them — not that the models are underperforming. The dashboard is how you find that out instead of guessing.

info

The dashboard is provisioned by Neurolabs and delivered outside the ZIA web app. Ask your Neurolabs contact for access, and for a walkthrough against your own catalogue — the panels are much easier to interpret with your own products in front of you. A version inside the web app is planned.

What it shows​

Variation health​

The working view of your catalogue at variation level. It lists your variations along with:

  • Variation detail, including channel lineage — where each reference image came from (Real Data Flow, validated results, synthetic generation, ZIA Capture)
  • Detection confidence, so you can see which variations the models are unsure about
  • Annotation confusion, particularly single versus multipack — where the models are mixing up pack formats of the same product

Channel lineage is the most useful column when diagnosing a problem. If a variation started performing badly, knowing that its last five references arrived from one particular source narrows the cause immediately.

Item health​

The same picture at item level, over time:

  • Version log — a history of changes to the item
  • Confidence before and after catalogue changes — did a change actually help?
  • Duplicate items — the same product existing more than once in your catalogue

Catalogue changes​

Compares performance before and after a catalogue change, across three measures:

MeasureWhat it means
AccuracyDerived from validations — did predictions get more correct?
ConfidenceThe margin between the chosen label and the next-best candidate. A wider margin means a less ambiguous decision
New neighboursWhether an item's five nearest neighbours have shifted. Movement here indicates the item's position relative to similar products has changed

Confidence is worth understanding properly, because it is a margin rather than a certainty score. A prediction can be correct but low-margin — meaning the runner-up was almost as close. Those are the predictions that flip to wrong as soon as shelf conditions change, so low-margin items are the ones worth fixing before they cause visible errors.

Catalogue duplicates​

Surfaces variations whose reference sets are too similar to tell apart, using cosine similarity across variation and seed type clusters.

Each pair carries an issue count: the number of times crops would have been mislabelled as the duplicate shown, based on the alternative predictions the models considered. This turns "these two look similar" into "this similarity has cost you N wrong labels", which is what lets you prioritise.

Identical seeds​

Flags reference images where the exact same embedding is used more than once — the same crop added twice, whether to one item or across several.

Identical seeds are pure downside. They consume slots against the 25-reference cap without adding any new visual information, and when the same crop sits under two different items they actively teach the models that two products are indistinguishable.

How to use it​

Diagnosing a product that is being confused​

  1. Find the affected variation in Variation health and check its detection confidence.
  2. Check Catalogue duplicates for that variation. A high issue count against another variation identifies the specific product it is losing to.
  3. Look at Identical seeds to rule out the same crop sitting under both.
  4. Inspect the reference images via channel lineage — a bad crop from one source is a common root cause.
  5. Fix the cause: remove the offending references, or add distinguishing ones via Accuracy Improvement.

Verifying that a change worked​

After approving new references or onboarding items, use Catalogue changes to check the change did what you intended. Look at confidence margins and the accuracy movement, not just the count of references added.

warning

Adding reference images can make accuracy worse. A crop that is genuinely the right product but visually ambiguous pulls two variations closer together. Always check the before-and-after rather than assuming more references is better — and if confidence dropped, the last batch of approvals is the first place to look.

Ongoing catalogue hygiene​

Used routinely rather than reactively, the dashboard supports a maintenance rhythm:

  • Duplicates and identical seeds — clear these first. They are unambiguous defects with a known cost, and removing them is the cheapest accuracy work available.
  • Low-confidence variations — the priority list for where new references are worth gathering.
  • Single vs multipack confusion — usually indicates a variation whose references contain the wrong pack format, rather than a model limitation.
  • Duplicate items — merge or remove, so that detections aren't split across two records of the same product.
tip

Work down the issue count, not the similarity score. Two variations can be extremely similar and cost you nothing if the products rarely appear together on shelf. The issue count reflects what actually went wrong in your images, so it is the better guide to where effort pays off.

Current limitations​

warning

The dashboard is at V1 and is actively evolving — panels and columns change between releases. Treat it as a diagnostic tool for investigation rather than a stable reporting surface, and don't build downstream reporting on its current shape.

Several related actions are scoped but not yet available: blocking duplicate reference images at upload, generating validation requests directly from the dashboard's problem variations, and automatically removing references identified as harmful. Today, the dashboard identifies the problem and the fix is applied through the normal catalogue and review workflows.