RAG that cites its sources
Retrieval is easy to demo and hard to trust. Here’s how we make enterprise search answers verifiable.
Read articleWhat it took to move an AI support suite from demo to resolving the majority of tickets on its own.
Our first prototype answered 90% of questions plausibly — and resolved almost none. Answering isn’t resolving. A resolved ticket means the refund was issued, the address changed, the customer satisfied.
The jump came from tools: order lookups, refunds within policy, account changes. Each tool has strict guards and an audit trail, so the agent can act without being trusted blindly.
We built an evaluation set from 2,000 real tickets and scored every change against it. Vibes don’t ship; numbers do.
Edge cases escalate with a full summary. Agents handle the routine; people handle the judgement calls — and both get better over time.
Have a question about this, or a project where it applies? Write to us — Slava reads every reply.
Retrieval is easy to demo and hard to trust. Here’s how we make enterprise search answers verifiable.
Read articleShipping LLM features without evaluation is shipping blind. How we build eval sets that actually catch regressions.
Read articleCaching, routing and smaller models — practical ways to make AI features affordable at scale.
Read article