Confidence-scored catalogue curation
Zapper Edge uses a model to identify products and fill catalogue attributes from raw records. Every result carries a confidence score, uncertain items go to a reviewer, and accuracy is tested continuously rather than assumed.
The challenge
Product catalogues are often built from raw installation or inventory records. Names are inconsistent, versions are ambiguous and lifecycle dates are missing, so identifying each product and keeping its attributes correct is slow, manual and hard to keep consistent.
A model can speed this up, but it can also guess. A confident wrong answer in a catalogue spreads into reports and decisions. Teams need results they can trace, a way to send doubtful cases to people, and proof the model has not drifted.
What it does
Product identification
A model reads raw signatures and proposes the product, so records with inconsistent naming are matched to a single catalogue entry instead of being interpreted by hand.
Vendor and version normalisation
Vendor names and versions are brought to a consistent form, so the same product is not listed several ways and version ambiguity is resolved or flagged.
Attribute enrichment
Missing catalogue attributes, including lifecycle dates, are filled using targeted web search, with the source recorded so each added value can be checked.
Confidence scoring
Every curated result carries a confidence score, so the pipeline knows which records can proceed and which need a person to look first.
Human review path
Low-confidence and edge-case records go to a reviewer. Their decisions are stored with the record and can feed later re-curation.
Continuous accuracy testing
Hallucination, edge-case, consistency and drift tests run against a known baseline, so accuracy is measured over time and regressions are caught early.
How it works
Pre-classify raw records
Raw signatures are sorted by input type first, so each kind of record follows the identification path that suits it.
Identify and enrich
The model proposes product, vendor and version, then enrichment adds missing attributes from search, with sources kept alongside each value.
Score and route
Each result is scored. Confident results proceed to the catalogue, and uncertain ones are routed to a reviewer with the evidence in front of them.
Test and learn
Reviewer feedback is stored, results are re-curated where needed, and regression and drift tests compare current output with the accuracy baseline.
Under the hood
- Models
- Any chosen LLM
- Prompted classification
- Confidence scoring
- Enrichment
- Web search APIs
- Source capture
- Normalisation rules
- Data
- Relational database
- Reviewer feedback store
- JSON
- Testing
- Ground-truth sets
- Regression tests
- Drift checks
Inside the harness
Evaluation
Hallucination, consistency, edge-case and drift tests run against ground truth, so accuracy is a measured figure you can track and not an assumption.
Guardrails
Results below the confidence threshold cannot enter the catalogue unreviewed, and enriched values must point to a recorded source.
Model routing
The identification step can use whichever approved model fits the task, and the model can change without redesigning the curation flow.
Audit
Each curated value is recorded with its score, source and any reviewer decision, so changes to the catalogue can be traced.
Business outcome
Less manual identification. The model handles routine matching and enrichment, so reviewers spend their time on genuine edge cases instead of every record.
Consistent catalogue data. Normalised vendor, product and version fields make the catalogue easier to search, report on and trust.
AI adopted under control. Confidence scoring and human review let teams use a model without letting unchecked guesses reach the catalogue.
Accuracy you can show. Baseline tracking and regression testing give evidence of how well the curation performs and when it changes.
Where it is used
- Matching raw software inventory records to a clean product catalogue with lifecycle information.
- Normalising supplier part descriptions into a single product master.
- Enriching a product list with missing attributes while recording where each value came from.
- Re-curating an existing catalogue after a model or rule change, with regression checks against a baseline.
Frequently asked questions
What happens when the model is not sure?
The record is routed to a reviewer with its evidence and score. Nothing below the confidence threshold enters the catalogue until a person approves it, and the decision is stored for later use.
How do you stop the model making things up?
Enriched values must carry a recorded source, and automated hallucination and consistency tests compare output with known-good examples. Failures are surfaced in testing, not found later in reports.
Can we choose the model?
Yes. The flow is vendor neutral, so a frontier, open-weight or small model can be used for each step, and the same tests show how a change affects accuracy.
Does it work for catalogues other than software?
The pattern of identify, normalise, enrich, score and review applies to any catalogue built from messy records. Fit is confirmed on your own sample data before a design is agreed.
See it on your own documents and data
Tell us what you want to automate and we will show you how it runs inside the harness.