Ecommerce ops · Production

leboncoin deploys LLM-powered Re-Ranker to improve search relevance and lift click and contact rates

The problem

leboncoin faces a complex search challenge: nearly 30 million monthly users and over 60 million classified ads each described in free-form user language create high volatility, while the search engine must respond within tens of milliseconds at thousands of requests per second, making relevant ranking technically demanding.

Workflow diagram · grounded in source
1
Ad embeddings pre-computed
ai_action
“compute a representation of all the ads available in our catalogue. Once all the ads in our catalogue are embedded in the vectors database (), we are ready for the real-time work”
2
User query triggers search
trigger
“the query is sent to ElasticSearch (ES), which first retrieves and ranks the ads using a TF-IDF like algorithm and custom functions”
3
ElasticSearch initial retrieval
integration
“a pool of ads is first retrieved and ranked by our ElasticSearch (with a TF-IDF like algorithm and custom functions). Only the top k of the pool of retrieved ads are sent to the Re-Ranker”
4
Re-Ranker scores top-k ads
ai_action
“The two encoders, made of a large language model (LLM) and custom layers are jointly trained to produce both an Ad and Query representation (a.k.a embedding or vector). Finally, the learned representations are concatenated and processed …”
5
Final ranking served to users
output
“those new Re-Ranker's scores are combined with the ElasticSearch's ones and the final sort takes place. The residual ads (which were not sent to the Re-Ranker) are simply appended after the top k that have been re-ranked to produce the f…”
Reported outcome

The Re-Ranker improved business targets (click and contact rate) by up to 5% and user experience KPIs (nDCG and average clicked and contacted positions) by up to 10%.

Reported metrics
Click and contact rateup to +5%
nDCG and average clicked and contacted positionsup to +10%
Reported stack
ElasticSearchLLMvectors database
Source
https://medium.com/leboncoin-tech-blog/serving-large-language-models-to-improve-search-relevance-at-leboncoin-2a364e5b6f76
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The Re-Ranker improved business targets (click and contact rate) by up to 5% and user experience KPIs (nDCG and average clicked and contacted positions) by up to 10%.

What tools did this team use?

ElasticSearch, LLM, vectors database.

What results were reported?

Click and contact rate: up to +5%; nDCG and average clicked and contacted positions: up to +10% (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Ad embeddings pre-computed → User query triggers search → ElasticSearch initial retrieval → Re-Ranker scores top-k ads → Final ranking served to users.

Related ecommerce ops cases