iFinder is an AI-powered enterprise search platform. It already had a language filter, but that filter assumed every document was written in one language. This case is about what happens when a document is written in two.
Enterprise Search · AI Behaviour
When
2025 · 3–6 months
My role
Product designer
Worked with
Head of Product, backend and frontend engineers, solution specialists, partner sales, customer success
01
01. The existing filter assumed every document had one language
Our clients work across several languages, and whole industries write in German while keeping their technical vocabulary in English. A document that was almost entirely German with a few English words counted as neither: filtering by German did not return it, and neither did filtering by English.
I validated the case with clients and shipped multilingual filtering. A document can now hold several languages, users select as many as they need, and results include everything matching any of them.
02
02. After shipping, the AI started behaving unexpectedly
Multilingual filtering only works if the model is consistent about which languages a document contains. Once it was live, search data and support requests pointed at the same thing: it was not.
The numbers said something was wrong but not what, so I read around 200 wrongly labelled documents by hand with the engineering team. Four kinds of document accounted for nearly all of them.
Read by hand, ~200 documents
The four kinds of document behind it
01
Very short documents
Adam Williams1 hour ago
I'll review it later.
>10 words: language UNKNOWN
02
Mostly figures and symbols
no readable words: UNKNOWN
03
Scanned pages
no extractable text: UNKNOWN
04
Two languages at once
Die neue Cloud-Infrastruktur ermöglich schnellerePerformanceund bessere skalierbarkeit für alleEnterpriseAnwendungen.
DEUTSCHENGLISHLanguage: UNKNOWN
Three options, built not described
Vibe coding let me put three working options on the table
Rather than argue about it in a document, I built each option with Claude and we compared them side by side. All three return the same documents — what changes is how much the interface tells you about them.
01No new UI
Leave the filter as it is
What it does
Check two languages, get anything matching at least one. Mixed documents come back, but look identical to the rest.
Why not
You still cannot ask for “only documents with both”, or tell which hit is the mixed one.
02Chosen
A switch for multilingual results
Off
One language or the other, and the filter says how many documents are being held back — 318.
On
Mixed documents join the results, labelled “Multilingual” with one tag per language.
What shipped
With the AI and engineering teams I set the rule for when the model cannot tell: if one language dominates and the unreadable part is small, the document takes the dominant one. Confidence indicators show when a guess is uncertain.
03Explored
Show each language’s share on hover
The idea
72% German with a French annex is a different document from 8% German. Scenario 2 labels both the same.
Why not now
Hover does not exist on touch, and it adds a second thing to read on a card people scan in seconds.
