Back office ops · Production

Dropbox implements semantic search with multilingual-e5-large, reducing empty search sessions by nearly 17%

The problem

Dropbox's Nautilus search engine relied on keyword matching that required users to recall exact file names or terms, missed contextually similar content, and could not serve multilingual users searching across languages.

Workflow diagram · grounded in source
1
Documents indexed as embeddings
ai_action
“Implementing semantic search at Dropbox scale would require mapping both new and existing documents to their embeddings and storing them in our indices”
2
User query entered
trigger
“The search begins when a user enters a query, which is converted into an embedding and compared against stored embeddings using vector search”
3
Query converted to embedding
ai_action
“semantic search operates by transforming user queries into embeddings and then performing vector search to retrieve results that align with the query's intent, rather than its literal terms”
4
Vector similarity search
ai_action
“Using nearest-neighbor algorithms, vector search identifies results based on meaning, ensuring that the retrieved content aligns with the user's intent”
5
Contextual results returned
output
“cross-lingual search. This advanced feature allows users to search in one language and receive relevant results in other languages”
Reported outcome

Semantic search powered by multilingual-e5-large delivered a nearly 17% reduction in empty search sessions and a 2% lift in search session success, and was made generally available to all Pro and Essential users in August 2024.

Reported metrics
empty search sessions reduction (ZRR)nearly 17%
search session success lift (qCTR)2%
Time spent searchingspend less time searching
Reported stack
Nautilusmultilingual-e5-largeMTEBKubeflow
Source
https://dropbox.tech/machine-learning/selecting-model-semantic-search-dropbox-ai
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Semantic search powered by multilingual-e5-large delivered a nearly 17% reduction in empty search sessions and a 2% lift in search session success, and was made generally available to all Pro and Essential users in Au…

What tools did this team use?

Nautilus, multilingual-e5-large, MTEB, Kubeflow.

What results were reported?

empty search sessions reduction (ZRR): nearly 17%; search session success lift (qCTR): 2%; Time spent searching: spend less time searching (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Documents indexed as embeddings → User query entered → Query converted to embedding → Vector similarity search → Contextual results returned.

Related back office ops cases