kovacIn regulated work, an answer without a source is not an answer. A reviewer needs to see the clause, the version of the procedure that applied on a given date, and the record of who changed the underlying documents. kovac is built so that every answer can be checked against the original.
Every answer cites the passage
Each answer lists its sources, and every citation opens the exact passage in the original document, shown next to the text kovac extracted from it. You can compare the original file with how kovac split it into searchable passages, page by page, to confirm nothing was lost or misread.
Long documents are handled as a whole. kovac keeps track of the section each passage belongs to and can return the surrounding section when a single paragraph is not enough context. Tables keep their rows, columns and headers, scanned documents are read with text recognition, and charts are described and, where possible, turned into tables.
The right version of the right document
Documents have versions, effective dates and an archive. You can record when a procedure comes into force and when it is superseded, and kovac answers from the version in force rather than whichever file was uploaded most recently. Metadata such as document type, owner or jurisdiction can be declared per collection, suggested automatically from the document text for a reviewer to accept, and used to filter answers. A question that names a document type or year can be narrowed automatically to the matching documents.
Access and audit
Access to each collection is granted per department, and access to each document can follow the permissions of the system it came from. The rule is enforced inside the database, so one team cannot receive answers from another team’s restricted documents.
Every change to knowledge is recorded in an audit log: uploads, deletions, metadata edits, permission changes, re-processing. You can see the full history of a single document and export the log to your own security monitoring. Staff sign in through single sign-on, and roles separate the people who manage documents from the people who only ask questions.
Measured, not assumed
Before you rely on the system, you can measure it. kovac generates candidate test questions from your documents, your experts accept or reject them, and the accepted set becomes the yardstick for retrieval quality. kovac can try several ways of splitting your documents into passages, measure each against that set, and let you apply the one that performs measurably better. Answers can also be scored automatically for faithfulness to their sources, and those scores are kept so a drop in quality is visible.
Skills that run code do so in an isolated sandbox, separate from the service that holds your data, and the credentials for model providers are stored encrypted.