How NovaCopilot works

Step by step

• The question is parsed: keywords, references such as act or case numbers, article numbers and question type.
• Everyday wording is mapped to the register's formal titles, so 'data protection act' and 'GDPR' land on the same act.
• The register is ranked with BM25 across title, reference, summary and keywords, with a boost for an explicit reference.
• The best-matching rows are quoted verbatim, with layer, reference and a path to the published node.
• The answer carries a receipt code, so a reader can verify afterwards which rows stood behind it.

What never happens

• No question leaves the device and no question is stored as text.
• No sentence is written freely: the short answer is a fixed template filled with the register's own data.
• No row is pulled from outside the platform's own register.

The limits

• An answer is a search in published sources, not legal advice.
• The register carries a read date per layer. A ruling issued after that date is not in the answer.
• When the register lacks the row, the assistant says so instead of guessing.

FAQ

Does NovaCopilot use a language model?

No. The answer is ranked and assembled from the register's own rows inside a fixed template. That makes it deterministic and traceable.

Is my question sent to a server?

No. The index is downloaded once to the device and the search runs locally, even offline.