Retail / Secondhand marketplace
Leboncoin: Ada, an internal knowledge search assistant
- Company
- Leboncoin
- Country
- France
- Adoption stage
- Discontinued
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Checked a source excerpt through a search tool
The work problem
In 2023 leboncoin saw an incident in which internal material leaked through a public LLM service. It wanted to try LLMs, but first it needed a controlled environment that protects its data and its teams.
Technology and data
At the end of 2023 a few machine learning engineers started an internal chatbot called Ada. They first tried putting Meta Llama 2 on AWS, but deployment was complex and conversation quality was poor. Anthropic Claude 2, accessed through AWS Bedrock, was better, so they changed direction. The legal and security teams confirmed the conditions that data does not leave AWS and is not used to retrain the model. From the middle of 2024 they built dedicated assistants connected to internal knowledge. Confluence and Lumapps documents were chunked, turned into embeddings and searched semantically in a Postgres vector database. Backstage was already indexed in OpenSearch, so its search API was used as it was. For policy documents and the org chart, the whole document was put into the model context.
Results
Adding an English keyword rewriter to the Backstage assistant lifted context relevance from 0.63 to 0.73 and returned the correct source link in 70 percent of cases. In the moderation assistant a rule based rewriter raised correctness by 10 percent, while an LLM based rewriter reached only 3 to 4 percent. However, a few engineers could not keep up with every team wanting its own assistant. Around the end of the first quarter of 2025 leboncoin retired Ada and decided to put its new internal assistant on the OpenAI platform.
Limits and open questions
The decision took the engineering and operating burden of self-hosting into account. The source does not disclose Ada's number of users or the amount saved.
Sources
Compiled from public sources. These are not results from ATF Works customers.