A reference build: we made this ourselves to show how we work. It is not a client engagement, there is no customer to name, and there are no measured outcomes to report - so this page does not pretend to have any.
A support assistant that cites where it got that
A reference build: a Greek-language assistant answering from a company's own documentation and past tickets, with a citation on every answer and a hard rule about what it does when it does not know.
- Client
- Reference build
- Sector
- Customer support
- Year
- 2026
- Built with
- Python, pgvector, Claude API, Viber API, Next.js
The numbers
- 2
- Numbers that cannot be gamed: how often it answered correctly with a citation, and how often it correctly refused
- 1
- Citation on every answer, so support can open the exact paragraph it used
- 2
- Channels answered, the website and Viber, because that is where Greek customers reply
The problem
The generic assistant a company bolts onto its website answers confidently in Greek and is wrong often enough that support ends up handling the same tickets plus the ones the assistant created. The failure is not the model's language ability - it is that it was never given the company's own material to answer from, so it answers from the internet's idea of a returns policy.
What we did
- 1
Retrieval over the company's own documents, product data and resolved tickets, with the source shown next to the answer. If a customer disputes what they were told, support can open the exact paragraph the assistant used.
- 2
Evaluated in Greek, on Greek tickets, because English performance does not transfer. The evaluation set was built from real resolved conversations and the model choice changed on the results - the obvious first pick was not the one that shipped.
- 3
Escalation is a designed outcome, not a fallback. Below a retrieval-confidence floor it stops, says it does not know, and hands the thread to a person with what it did find attached, so the human does not start from nothing.
- 4
It answers on Viber as well as the website, because that is where Greek customers actually reply. An assistant that only lives in a web widget is answering the wrong channel.
Where it got to
The measurable thing is not deflection rate, which any assistant can inflate by refusing to escalate. It is the pair of numbers worth arguing about: how often an answer was correct and cited, and how often it correctly refused. Both are measured on the company's own tickets before anything is switched on for customers.
A worked example
Deflection rate is the figure to distrust: any assistant inflates it by refusing to escalate. The pair that cannot be:
- Answered, correct, citation works
- measured on yours
- Refused when it should have
- measured on yours
- Evaluation set
- your resolved tickets
- Language it is scored in
- Greek
No accuracy figure appears here because we have not run it on your data, and a number from someone else's data would tell you nothing.