Every answer carries the document, the page and the quotation
Lotte E&C · commercial deployment
Citation attribution on specification Q&A: 97.8%, measured in the Lotte E&C production environment. The same engine has been proven in two further domains — 92.3% search accuracy on construction drawings and contracts (KT Estate, in operation) and 99% symbol recognition on fire protection drawings (Seoul Metropolitan Fire & Disaster HQ, pilot). The three figures are measured in different environments on different tasks and are not comparable with one another.
basis: measured on specification Q&A in the Lotte E&C production environment
Context — background
When people query internal knowledge and technical documents, an answer without evidence is worse than no answer at all — above all where the answer feeds a working decision, as it does with specifications, standards and codes.
The existing workflow and its bottleneck
A general-purpose chatbot answers without a source, so the engineer had to go back to the original document and verify the answer anyway.
Data supplied
A corpus of internal specifications, technical documents, standards and codes.
doAZ capabilities used
Retrieval-augmented generation with enforced source attribution — every answer carries the document, the page and the quotation, and when the evidence is weak the system abstains before anyone has to ask.
What the AI does
Understanding the question, retrieving the evidence, and drafting an answer with its citations attached.
What rules, calculation and verification do
Citation verification — an attribution-checking layer that confirms the answer actually corresponds to the source text.
Where a person reviews
The answer is reference material. Working decisions and official calculations are made by a person, on the basis of the source document.
Measured outcome and its conditions
Citation attribution rate 97.8%
Each figure is measured in its own environment: Lotte E&C in commercial deployment, KT Estate in operation, and the Seoul Metropolitan Fire & Disaster HQ pilot.
Known limitations
The attribution rate is measured against each customer's environment and corpus; porting it to a new domain requires re-measurement. An LLM answer is not an official calculation result.
Customer approval scope
Customer names and figures approved for publication — including a jointly filed patent application (Lotte E&C).
Other cases
We agree on the accuracy definition against a frozen gold set and then measure it — a structure in which cherry-picking is impossible.
Request a technical verification demo