DigUp: local AI search should lead to the right evidence

“We discussed this somewhere, but I cannot find the file.” For many teams, that costs more time than writing another paragraph with AI. Improving retrieval may be a better first step than adding another assistant that generates answers.

Jirka Herník included DigUp in Mastermind 45, published on October 22, 2026. This dates the video, not the application’s original release.

DigUp’s README describes local content search across selected files, including documents and recordings. Windows and Linux are marked beta. Local-processing claims come from the project; I have not audited its network traffic.

Start from the project’s official release. Version 0.6.2 is dated October 21 and adds spreadsheet and presentation reading while addressing repeated-indexing issues. After installation, select only the test folder and let indexing finish. This earlier release provides context for the new practical recommendation.

Meaning helps, but the source decides

Its foundation is Google’s EmbeddingGemma 2, which maps text, images, audio and video into a shared vector space. Semantic similarity can locate material without an exact wording match.

The useful outcome is opening and checking the correct source. A similar document may be outdated, and a mention of a decision does not establish that it was approved.

Pilot one closed project

I would start with a copy of an approved folder, not the entire disk. Prepare twenty questions with known source files: the final brief, a deadline decision, a technical appendix or a recording segment assigning responsibility.

Add an older version of a document and three questions with no answer in the folder. Judge retrieval against the prepared answer key rather than whether a result sounds plausible.

Measure time to the correct source, version confusion and initial indexing time. Failed retrieval must remain visible instead of encouraging someone to select a weak match.

Fit it into existing work

Sales could find evidence for a CRM brief, support an original technical note, and a content team approved footage. In the first pilot, a person should confirm the source before forwarding its link. This article does not claim a native n8n or Make connector.

Local execution does not settle access control. Who can access the computer, index and backups? What happens when a file leaves an approved folder? Test these before using production records.

Evaluate DigUp by the cost of finding and verifying the right evidence. If it saves search time without increasing misidentification, it can be useful before a complex company-wide RAG deployment. This is a proposed pilot, not a benchmark I have run.

The cover is a photorealistic AI illustration, not DigUp’s actual interface.