RAG over 220k internal docs — accurate enough to ship to customers
Mailaxis · AI · SaaS
220k
Documents indexed
91%
Answer accuracy (eval suite)
13 wk
Delivery timeline
The challenge
Mailaxis had 220,000 internal support and product documents accumulated over eight years. Customers couldn't find answers. Support agents couldn't find answers.
Our approach
We built a production RAG pipeline with a custom chunking strategy, hybrid semantic + keyword retrieval, and a rigorous evaluation suite testing accuracy against 400 golden questions before every deploy.
The outcome
The feature launched with 91% accuracy. Support escalations from it are under 9% — below the team's 12% target.
“Knowing our accuracy before we ship is genuinely new for us. The eval suite is something I didn't expect to value as much as I do.”
— James Osei, VP Product, Mailaxis