Quality assurance · Production

Zalando builds LLM-as-a-judge search quality assurance framework for multi-market launches

The problem

Zalando's pre-launch search quality assurance relied entirely on human experts manually sampling and translating queries, annotating errors, and diagnosing root causes. The process was not scalable and was reactive by nature — issues were only caught after launch when real-user signals such as CTR existed. For entirely new markets, those signals did not exist at all.

Workflow diagram · grounded in source
1
Historical query clustering by NER
ai_action
“we have a Named entity recognition (NER) engine to extract a wide range of attributes from the search queries, such as product name, brand, colour, size, season, occassion, material, etc. If we group search queries tagged by the same set…”
2
LLM query translation
ai_action
“We can translate these queries to other languages using an LLM. This enables us to reuse search scenarios from existing markets for new markets with different languages, while having translated scenarios keep the same search intents.”
3
Search result retrieval
integration
“This pipeline step spawns a task in our Kubernetes cluster, runs through the test queries and submits them to the search microservice to retrieve the search results. The results are then kept in an in-memory cache for the evaluation step.”
4
LLM-as-a-judge relevance scoring
ai_action
“The search results are then evaluated by LLM-as-a-judge, which we used GPT-4o during pre-market launch process. The pipeline submits the search results, their product data and images to the LLM and get the relevance scores for each resul…”
5
NER tag consistency validation
validation
“each search query is processed by the NER engine to extract its NER tag attributes. This allows us to compare the NER tags of the original search query and the translated search query, and identify inconsistencies that can lead to search…”
6
Issue flagging and report output
output
“we could easily identify segments that did not perform well”
7
Engineer investigation and fix
human_review
“engineers could investigate further and identify that NER engine may have got a lemmatization issue for terms related to "desporto", "desportivo", "desportiva"”
Reported outcome

The LLM-as-a-judge evaluation framework identified multiple NER and search quality issues in Portuguese and Greek markets before go-live, enabling engineers to fix them pre-launch.
A full run covers 1,500 search segments with 25 results each, completes in 3-5 hours, and costs around 250 USD — compared to days of human evaluation.

Reported metrics
Cost per evaluation runaround 250 USD
Search segments evaluated per run1,500
Search results evaluated per query25
Evaluation run duration3-5 hours
Show all 6 reported metrics
cost per evaluation runaround 250 USD
search segments evaluated per run1,500
search results evaluated per query25
evaluation run duration3-5 hours
human evaluation time comparisonwould take days
LLM-human judgement correlationhigh correlation with human judgement
Reported stack
GPT-4oApache AirflowKubernetesNakadiElasticacheNER
Source
https://engineering.zalando.com/posts/2026/03/search-quality-assurance-with-llm-judge.html
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The LLM-as-a-judge evaluation framework identified multiple NER and search quality issues in Portuguese and Greek markets before go-live, enabling engineers to fix them pre-launch.

What tools did this team use?

GPT-4o, Apache Airflow, Kubernetes, Nakadi, Elasticache, NER.

What results were reported?

Cost per evaluation run: around 250 USD; Search segments evaluated per run: 1,500; Search results evaluated per query: 25; Evaluation run duration: 3-5 hours (source-reported, not independently verified).

How is this quality assurance AI workflow structured?

Historical query clustering by NER → LLM query translation → Search result retrieval → LLM-as-a-judge relevance scoring → NER tag consistency validation → Issue flagging and report output → Engineer investigation and fix.