OCR document search across scans, PDFs and metadata

Optical character recognition runs on uploaded images and on PDFs that contain only a scanned page image rather than a text layer. The recognised text is written into the search index and attached to the document, so a signed delivery note photographed by a driver becomes findable by the consignment number printed on it.

This is the difference between a digital archive and a searchable one. Scanning a filing cabinet into image PDFs produces something you can store but not use; running OCR over it produces something you can answer questions from. Recognition accuracy depends on the source — a clean 300 dpi scan reads far better than a photograph taken at an angle in poor light — so the honest expectation is that OCR makes most of your archive searchable, not all of it.

Global search queries the extracted text and the structured metadata together. That matters because the useful question is almost never purely one or the other: "the supplier agreement with a renewal date this quarter that mentions indexation" needs the date from a metadata field and the word from the document body, and running two searches and intersecting them by hand is how people give up.

Results are permission-filtered before they are returned. A user never sees a result they are not entitled to open, and the search itself does not leak the existence of documents through result counts.

Metadata fields are defined per document type, so a contract carries counterparty, value, term and renewal date while a policy carries owner, review date and approval reference. Mandatory fields can be enforced at upload, which is the only reliable way to keep a metadata model from decaying into blanks.

Where AI extraction is enabled, those fields are populated from the document content and presented for confirmation rather than silently accepted, so the metadata stays trustworthy enough to report from.

Filters apply on top of a query instead of replacing it: category, folder, document type, owner, date range, approval status, retention class and any custom metadata field. The behaviour people expect from an online shop is the behaviour that works here too — refine, look, refine again.

A filtered search can be saved and shared, which turns an ad-hoc query into a working view for a team. It is also how disclosure exercises stay reproducible: the same saved search run a month later returns the same set plus anything new that matches.

Keyword search fails in a specific and frustrating way: you remember what a document was about but not what it said. Semantic search matches meaning, so a query for "the lease that renews in the third quarter" can return the right agreement even though it uses the phrase "term expires 30 September".

Semantic and keyword results are presented together rather than as separate modes, because forcing a user to choose a search algorithm is asking them to solve our problem.